Paper deep dive
Task-Driven 3D Printability Assistance via Geometry- and Knowledge-Grounded LLM Reasoning
Zhaoda Du, Qiaojie Zheng, Xiaoli Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/25/2026, 8:15:17 AM
Summary
This paper introduces a task-driven 3D printability assistance framework that leverages Large Language Models (LLMs) grounded in geometry evidence and a structured Task-Material-Process knowledge graph. The system analyzes STL models and natural language task descriptions to generate pre-print recommendations, including material selection, process parameters, and design guidance. Evaluated on 96 physical trials, the method achieved 75.0% printability and 88.9% task suitability, significantly improving material selection accuracy over pure LLM baselines.
Entities (9)
Relation Signals (6)
Task-Driven 3D Printability Assistance Framework → uses → Large Language Models
confidence 98% · this paper leverages the reasoning and language-understanding capabilities of large language models (LLMs)
Task-Driven 3D Printability Assistance Framework → uses → Task-Material-Process Knowledge Graph
confidence 97% · combines this geometry-grounded evidence with a structured Task-Material-Process knowledge graph
Task-Driven 3D Printability Assistance Framework → achieves → Printability
confidence 95% · The proposed method achieves 75.0% printability over 96 physical validation trials
Task-Driven 3D Printability Assistance Framework → achieves → Task Suitability
confidence 95% · with 88.9% task suitability among successfully printed samples.
Geometry-Grounded Layer → partof → Task-Driven 3D Printability Assistance Framework
confidence 95% · The proposed framework consists of three functional layers... The Geometry-Grounded Layer converts the uploaded STL geometry
Task-Driven 3D Printability Assistance Framework → improves → Gemini 2.5 Flash-Lite
confidence 90% · It also improves Gemini 2.5 Flash-Lite material-selection accuracy from 37.5% under pure LLM to 90.0%.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Printability assessment in additive manufacturing is typically conducted at the geometry level before printing to determine whether a computer-aided design (CAD) model or stereolithography (STL) file can be successfully fabricated. Task suitability, in contrast, is usually evaluated after printing to determine whether the fabricated part satisfies the requirements of its intended use. As a result, for non-expert users to print functional parts, unsuitable material or process choices may only be identified after fabrication, leading to repeated printing, material waste, and user frustration. To address this challenge, this paper leverages the reasoning and language-understanding capabilities of large language models (LLMs), while grounding the reasoning with geometry evidence and structured material/printer knowledge to generate reliable pre-print recommendations. Given a stereolithography (STL) model and a natural-language task description, the framework generates a structured recommendation covering printability, material choice, process parameters, design guidance, risks, and explanations. We evaluate the framework on focused STL benchmark scenarios with novice-style task descriptions. The proposed method achieves 75.0% printability over 96 physical validation trials, with 88.9% task suitability among successfully printed samples. It also improves Gemini 2.5 Flash-Lite material-selection accuracy from 37.5% under pure LLM to 90.0%. Expert evaluation further shows improved report quality, while post-print feedback improves recommendations on selected problematic cases. These results suggest that user task intent, geometry evidence, and structured material knowledge are all important for reliable task-driven printability assistance.
Tags
Links
- Source: https://arxiv.org/abs/2608.22128v1
- Canonical: https://arxiv.org/abs/2608.22128v1
Trouble viewing inline? Open PDF directly →
Full Text
70,679 characters extracted from source content.
Expand or collapse full text
Task-Driven 3D Printability Assistance via Geometry- and Knowledge-Grounded LLM Reasoning Zhaoda Du a , Qiaojie Zheng a and Xiaoli Zhang a,∗ a Colorado School of Mines, 1500 Illinois St., Golden, 80401, CO, USA A R T I C L E I N F O Keywords: Additive manufacturing 3D printability Large language models Knowledge graph A B S T R A C T Printability assessment in additive manufacturing is typically conducted at the geometry level before printing to determine whether a computer-aided design (CAD) model or stereolithography (STL) file can be successfully fabricated. Task suitability, in contrast, is usually evaluated after printing to determine whether the fabricated part satisfies the requirements of its intended use. As a result, for non-expert users to print functional parts, unsuitable material or process choices may only be identified after fabrication, leading to repeated printing, material waste, and user frustration. To address this challenge, this paper leverages the reasoning and language-understanding capabilities of large language models (LLMs), while grounding the reasoning with geometry evidence and structured material/printer knowledge to generate reliable pre-print recommendations. Given a stereolithography (STL) model and a natural-language task description, the framework generates a structured recom- mendation covering printability, material choice, process parameters, design guidance, risks, and explanations. We evaluate the framework on focused STL benchmark scenarios with novice-style task descriptions. The proposed method achieves 75.0% printability over 96 physical validation trials, with 88.9% task suitability among successfully printed samples. It also improves Gemini 2.5 Flash- Lite material-selection accuracy from 37.5% under pure LLM to 90.0%. Expert evaluation further shows improved report quality, while post-print feedback improves recommendations on selected problematic cases. These results suggest that user task intent, geometry evidence, and structured material knowledge are all important for reliable task-driven printability assistance. 1. Introduction Assessing printability before fabrication is important in Additive Manufacturing (AM) for time and cost reduction. In current 3D printing workflows, printability is typically un- derstood at the geometry level as whether an input computer- aided design (CAD) or stereolithography (STL) model can be fabricated under a given printer, material, and process setup without major appearance failures. However, as AM has become increasingly mature, it is no longer limited to hobbyist prototyping but is now widely used to fabricate replacement parts, custom fixtures, and other application- dependent objects. In these applications, geometric printabil- ity alone is insufficient, because a printed part must also satisfy the requirements of its intended use. Simultaneously considering geometric printability and task suitability is challenging because the two objectives operate at different levels and are not independent [13, 2]. Geometric printability concerns whether the input geometry can be fabricated reliably under a given setup [12], whereas task suitability concerns whether the printed part can satisfy the requirements of its intended use after fabrication. At the same time, these objectives are often coupled through material and process choices [4]: a material or setting that better supports the intended use may introduce new fabri- cation difficulties, while a setup that improves geometric printability may reduce suitability for the target application [36, 2, 13]. ∗ Corresponding author. zhaoda_du@mines.edu (Z. Du); zheng@mines.edu (Q. Zheng); xlzhang@mines.edu (X. Zhang) This challenge is not well addressed by existing au- tomated printability assessment methods [12, 25, 29, 35]. Although these methods can identify geometry-related risks such as overhangs, wall-thickness issues, or mesh defects, they do not infer the user’s intended use or jointly recom- mend material and process choices under task requirements. For example, a geometry-level checker may judge a model as printable with PLA, but the printed part may soften or deform if the intended task requires use under heat exposure. Task suitability has often been addressed through expert empirical initial material/process selection combined with post-print evaluation and iteration [32, 10]. Consequently, task suitability is highly dependent on expert knowledge and often requires repeated printing before a suitable part is obtained. Although prior work has recognized printability and task suitability as related but distinct objectives [2], exist- ing studies do not jointly reason about them for a specific input geometry and user task before printing. In practice, these two aspects are often handled sequentially rather than jointly, as illustrated in Figure 1. This workflow is not fully automated and usually requires expert intervention to select materials, adjust process parameters, interpret printability risks, and revise the design after failures. As a result, it is inefficient, failure-prone, and highly dependent on expert knowledge [32, 10]. To address this joint reasoning gap, we leverage the reasoning and language-understanding capabilities of LLMs to connect ambiguous user task intent with geometry, material, and process requirements. Specifically, we develop a geometry- and knowledge-grounded LLM Du et al.: Preprint submitted to ElsevierPage 1 of 16 arXiv:2608.22128v1 [cs.AI] 22 Aug 2026 Input: Geometry Input: Geometry Rule-based / data-driven User: Define material, printer, process parameter Printability Print •Task-unaware •Expert knowledge required •High failure risk increases trial-and-error iterations. Conventional Methods Task-Driven Printability Assistance LLM-guided reasoning Printability Print Task suitability User: Task “I want to print a soft toy” Our Task-Driven Framework To user: •Material/ Process/Geometry Suggestions •Explanation •Task-aware •Expert knowledge not required •Low failure risk supports first-pass success. Task suitability Figure 1: Comparison between conventional printability evaluation and the proposed task-driven printability assistance framework. framework for task-driven printability assistance, as shown in Figure 2. Given a natural-language task intent and an input STL model, the framework first converts the geometry into a structured evidence package through rule-based mesh-level computation, orientation scoring, heuristic risk generation, parameter baselines, and material candidates. It then combines this geometry-grounded evidence with a structured Task-Material-Process knowledge graph and uses constrained multi-objective LLM reasoning to generate a structured recommendation report. The report includes an overall printability assessment, task-relevant risks, material and process-parameter suggestions, design guidance, and explanations. An optional post-print self-improvement module records observed outcomes and user ratings to update memory, heuristics, and future scoring. The main contributions of this work are summarized as follows: 1. A task-driven formulation of printability assis- tance. We formulate printability assistance as a task- dependent problem, where the same geometry may require different material, parameter, design, and risk recommendations depending on the user’s intended task. 2. A geometry- and knowledge-grounded LLM framework for task-driven printability assistance. We propose a framework that combines geometry evidence, a structured Task-Material-Process knowledge graph, and constrained LLM reasoning to generate task-driven printability reports with risks, explanations, and material/parameter/design suggestions. 3. Evaluation methodology for task-driven printabil- ity assistance. We evaluate the framework across mul- tiple LLM backbones using novice-style task descrip- tions and same-geometry different-task cases, with evaluation organized around geometry-level printabil- ity evaluation, task suitability evaluation, material se- lection accuracy, and expert evaluation of report qual- ity. We further assess the self-improvement module on selected problematic cases. 2. Related Work 2.1. Rule-Based Printability Assessment Traditional printability assessment methods often eval- uate whether a CAD or STL model satisfies geometry- or process-specific constraints under a predefined manufac- turing setup. Automated printability checkers typically ex- tract geometric features and compare them with printer constraints such as build volume, minimum feature size, wall thickness, overhang limits, support requirements, and mesh validity. For example, Lu formulated printability checking as a rule-based verification problem over object features and printer profiles [25], while Fudos et al. characterized 3D printability through geometry- and technology-dependent printability scores [12]. Related work on support generation, model decomposition, build-orientation planning, and multi- axis support-free fabrication further shows that printability depends on orientation, support strategy, and process plan- ning rather than geometry alone [21, 37, 26, 17, 23, 15]. These methods provide interpretable and reproducible manufacturability evidence, which is important for practical Du et al.: Preprint submitted to ElsevierPage 2 of 16 User Inputs Task Intent “I want to print a soft toy” Geometry(STL) Geometry-Grounded Layer(Rule-Based) Mesh-level Computation •Triangle metrics •Bounding box •Volume estimation •Topology checks •Watertightness •boundary edges •non-manifold edges •degenerate triangles (Z+)(Z-)(X+)(X-)(Y+)(Y-) Orientation Scoring(lower better) 푆 표 = 0.58푅 푠푢푝표푟푡 + 0.16min1,푅 푠푙푒푛푑푒푟 /10 + 0.121−min(1,푅 푏푎푠푒 )+ 0.14푅 ℎ푒푖푔ℎ푡 Heuristic Evidence and Scores •Heuristic Risks 푅: support, overhang, adhesion, slenderness, and topology risks. •Design Suggestions 퐷: rule-triggered CAD modifications to reduce printability risks. •Parameter Baseline Θ 0 :initial slicer settings for orientation, layer height, support, adhesion, speed, and cooling. •Material Candidates 푀 cand :keyword-based material matching. •Printability Scores 푃: heuristic scores for current and optimized orientations. •Best Orientation 표 ∗ :orientation candidate with the lowest estimated printability risk. Geometry- and Knowledge-Grounded LLM Reasoning Layer Task-Material-Process Knowledge Graph Constrained Multi-objective Reasoning 1.Understand task intent 2.Analyze geometry evidence and risks 3.Retrieve relevant material/printer knowledge according to knowledge graph 4.Compare material candidates: 5. Balance printability and task alignment 6. Generate final recommendations Reasoning process: Post-Print Self-Improvement Module(optional) •Success/Failure •Issues •User rating Output (Structured) •Add new failure cases •Refine heuristics & rules •Improve retrieval & scoring Update Memory & Heuristic Priors Select 표 ∗ =argmin푆 표 Figure 2: Overview of the proposed task-driven printability assistance framework. decision support. However, they are generally setup-first: the printer, material, orientation objective, and process con- straints are assumed before assessment. As a result, they can identify geometry-related risks, but they usually do not infer the user’s task intent or adapt material, parameter, and design recommendations to task suitability requirements. 2.2. Data-Driven Printability Prediction Data-driven methods extend rule-based assessment by learning relationships between geometry, process conditions, and print outcomes. Geometry-centered approaches use su- pervised machine-learning models, decision-tree methods, or neural networks to learn from CAD features, local de- scriptors, STL-derived geometric representations, or printed test artifacts [29, 35, 18]. For example, Trovato and Cicconi used a decision-tree approach for early evaluation of 3D models [35], while Henn et al. trained a convolutional neural network to detect geometry-related printability issues from STL-derived inputs before printing [18]. Data-driven process-aware models further incorporate material behavior and process parameters. Hammoud et al. combined experimental characterization with machine- learning prediction for 3D concrete printing, including geometric-quality prediction, inverse parameter selection, and printability maps [16]. Other studies model geometric deviation under different printhead and process conditions or use process-window prediction and in-situ defect monitoring to account for process context and observed outcomes [22, 6, 14, 33, 11]. Although data-driven and process-aware models can provide useful predictions and, in some cases, parameter- selection guidance, they remain limited for task-driven assistance. They often require substantial training data, provide limited explanation for decisions, and are tied to Du et al.: Preprint submitted to ElsevierPage 3 of 16 specific machines, materials, process windows, geometry families, or quality metrics. 2.3. LLM-Based Manufacturing Assistance Large language models(LLMs) offer a useful interface for manufacturing assistance because they can interpret natural-language instructions, reason over ambiguous goals, and generate explanations understandable to non-expert users [30, 24, 8]. Recent studies have explored LLMs and vision-language models for additive-manufacturing knowledge querying, defect prediction, process monitoring, and manufacturing decision support [41, 40, 19, 30, 31, 9, 5]. AMGPT investigates LLM-based contextual querying in additive manufacturing [5], while FDM-Bench evaluates LLM competence on domain-specific additive- manufacturing tasks [9]. AdditiveLLM studies LLM-based defect prediction in metal additive manufacturing [31], and LLM-3D Print demonstrates that language-model agents can monitor print images, diagnose defects, and suggest or execute corrective actions during material-extrusion printing [19]. Broader reviews also discuss the integration of LLMs into digital manufacturing workflows [30]. These studies show that LLMs can provide flexible inter- action, technical-context interpretation, and human-readable explanations. However, existing LLM-based manufacturing systems focus on knowledge querying, monitoring, or defect diagnosis rather than pre-print task-conditioned assistance from an uploaded STL model. 3. Methodology The proposed framework consists of three functional layers and an optional post-print self-improvement module, as shown in Figure 2. The Geometry-Grounded Layer models the task-geometry relation by combining geometry/topology analysis with task cues and expert- defined fabrication heuristics. The Task-Material-Process knowledge graph provides structured task-material- process relations by linking task requirements, material properties, and printer/process constraints. The Geometry- and Knowledge-Grounded LLM Reasoning Layer then integrates these task-geometry and task-material-process relations to perform task-driven reasoning over geometry, material, and process decisions. In this way, the framework generates recommendations that are not only geometrically printable but also aligned with the user’s intended task. The optional post-print self-improvement module supports post-print refinement using observed outcomes and user feedback. 3.1. Geometry-Grounded Layer The Geometry-Grounded Layer converts the uploaded STL geometry 퐺 and lightweight task cues from 푇 into a structured geometry-grounded evidence package 퐸 G for downstream LLM reasoning. This layer consists of three steps: mesh-level computation, orientation scoring, and heuristic evidence generation and scoring. Its output is defined as: 퐸 퐺 = geo (퐺,푇) = 퐹 mesh ,푄 topo ,푆 표 ,푅,퐷,Θ 0 ,푀 0 cand ,표 ∗ . (1) Here, 퐹 mesh and 푄 topo denote mesh and topology evidence, 푆 표 denotes orientation-dependent printability scores, 푅 de- notes heuristic risks, 퐷 denotes design hints, Θ 0 denotes an initial parameter baseline, 푀 0 cand denotes keyword-based initial material candidates, and 표 ∗ denotes the selected best orientation. The following subsections describe how these components are generated. 3.1.1. Mesh-level Computation The mesh-level computation step generates the mesh and topology evidence (퐹 mesh ,푄 topo ). The uploaded STL is parsed as a triangle mesh, from which the layer computes geometric quantities such as model dimensions, surface area, and estimated volume. It also checks topology-related indi- cators such as watertightness, boundary edges, non-manifold edges and degenerate triangles. These outputs provide low- level evidence about whether the uploaded geometry is valid and suitable for slicing [25, 12]. 3.1.2. Orientation Scoring The orientation scoring step generates the orientation- dependent score푆 표 and the selected orientation candidate표 ∗ . To account for orientation-dependent printability, the layer evaluates six canonical build orientations: = 푧 + ,푧 − ,푥 + ,푥 − ,푦 + ,푦 − .(2) For each orientation, the layer computes support-risk, base-contact, slenderness, and height-dominance indicators. These indicators are combined into a heuristic orientation score: 푆 표 =0.58푅 support + 0.16min ( 1, 푅 slender 10 ) + 0.12 ( 1 − min(1,푅 base ) ) + 0.14푅 height . (3) The weighting coefficients in Eq. (3) were empirically determined through preliminary experiments on a validation set of primitive geometries to balance the relative printability impacts of support burden, slenderness-induced instability, bed adhesion risk, and height dominance [10, 34]. The four orientation-dependent indicators are defined as 푅 support = 퐴 support 퐴 total , 푅 base = 퐴 base 퐴 footprint , 푅 slender = 퐻 max(min(푆 푥 ,푆 푦 ),0.01) , 푅 height = 퐻 max(푆 푥 ,푆 푦 ,퐻,0.01) . (4) Here, 퐴 support is the area of downward-facing near- horizontal surfaces, 퐴 total is the total surface area, 퐴 base is the estimated build-plate contact area, 퐴 footprint is the Du et al.: Preprint submitted to ElsevierPage 4 of 16 (a)(b) Figure 3: Partial visualization and task-conditioned evidence tracing of the Task-Material-Process Knowledge Graph. (a) The visualization shows the main node and edge types in the KG, including application nodes, material-property nodes, candidate- material nodes, and printer/process constraint nodes. These nodes organize task-related application contexts, material properties, material options, and printing-feasibility constraints. (b) Given a user task, the framework traces task-relevant application and property evidence through the KG to candidate materials. In this example, the task emphasizes structural use and heat resistant. PC receives the strongest converging evidence. footprint bounding-box area, 퐻 is the build height, and 푆 푥 ,푆 푦 are the horizontal spans in the build-plate plane. The orientation with the lowest score is selected as the best orientation candidate: 표 ∗ = argmin 표∈ 푆 표 .(5) The resulting 푆 표 and 표 ∗ are included in the geometry- grounded evidence package and passed to the downstream heuristic evidence generation and LLM reasoning steps. 3.1.3. Heuristic Evidence and Scores The heuristic evidence generation step converts mesh evidence, topology evidence, orientation indicators, and lightweight task cues into rule-triggered fabrication evidence (푅,퐷,Θ 0 ,푀 0 cand ). High support risk triggers support-related warnings, weak base contact triggers adhesion guidance, invalid topology triggers mesh-repair warnings, and high slenderness triggers stability warnings. Lightweight task cues further influence which risks and suggestions are emphasized. For example, clean-appearance tasks emphasize support scarring and underside quality, fit-related tasks emphasize dimensional accuracy and stability, and heat-related tasks emphasize material and shape-retention concerns. Based on these rule triggers, the layer also produces CAD-facing design hints 퐷, initial slicer-parameter baselines Θ 0 , and keyword-based initial material candidates 푀 0 cand [1, 20]. 3.2. Task-Material-Process Knowledge Graph The partial KG visualization in Figure 3 shows the main node and edge types, including application nodes, material-property nodes, printer/process constraint nodes, and candidate-material nodes. The complete Task-Material- Process KG contains 52 typed nodes, including 8 application nodes, 12 material-property nodes, 7 printer/process constraint nodes, and 25 candidate-material nodes. It also contains 347 typed edges, including 72 application-material edges, 96 property-material edges, and 179 printer/material feasibility edges. Application and material-property nodes provide the task-side semantic structure, representing application contexts and task-relevant material requirements. Candidate-material nodes encode available material options, while printer/process constraint nodes encode feasibility and process-related constraints under the printing setup. Together, these nodes provide traceable task-material- process knowledge for downstream LLM reasoning. The Task-Material-Process KG was constructed through an expert-supervised, documentation-grounded, and LLM- assisted process. We first defined the graph schema, includ- ing the allowed node types, edge types, candidate-material scope, printer/process constraints, and scoring terms. Under this fixed schema, the LLM was used as an assistant to organize information [3] from official manufacturer material- printer documentation, group application and property cat- egories, suggest missing task-property-material links, and initialize heuristic evidence strengths. All suggested nodes, links, and weights were normalized and checked by the ex- perts before being used for retrieval and scoring. Therefore, the KG is not an unconstrained LLM-generated knowledge base; it is a schema-constrained material-printer knowledge layer used to provide traceable evidence for material recom- mendation. Table 1 summarizes which parts of the KG were expert-defined and which parts were LLM-assisted. Du et al.: Preprint submitted to ElsevierPage 5 of 16 Table 1 Responsibility split in Task-Material-Process KG construction. KG componentExpert-definedLLM-assisted Graph schemaDefined node types, edge types, and allowed relation types No schema change Candidate material scope Fixed the material set used in the experiments Did not introduce materials outside the allowed scope Printer/process scopeDefined printer-feasibility fields and process- constraint categories Helped organize compatibility information from documentation Application nodesDefined the application-node structureHelped organize documentation-derived appli- cation categories and synonyms Property nodesDefined the material-property node structure Helped extract and group material-property descriptions Material nodesFixed candidate material identities and naming conventions Helped normalize surface forms and synonyms EdgesDefined allowed edge meanings, such as application-material, property-material, and printer-material feasibility Suggested missing task-property-material links under the fixed schema Edge weightsDefined scoring terms and normalization rules Suggested initial heuristic evidence strengths 3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer The geometry- and knowledge-grounded LLM reason- ing layer is the final task-driven reasoning layer of the frame- work. It receives the user task description 푇 , the geometry- grounded evidence package 퐸 퐺 , and the KG-based material evidence package 퐸 퐾퐺 , and generates the final structured recommendation report: 푂 = LLM (푇,퐸 퐺 ,퐸 퐾퐺 ) = 푦,푀,푅 푇 ,Θ,퐷 푇 ,퐸. (6) where 푦 is the overall verdict, 푀 is the final material recom- mendation, 푅 푇 is the task-relevant risk set, Θ is the process- parameter guidance, 퐷 푇 is the task-conditioned design guid- ance, and 퐸 is the rationale explaining the recommendation. This structure makes the output auditable because each com- ponent can be traced back to task intent, geometry-grounded evidence, or graph-retrieved material/process knowledge. The geometry-grounded evidence 퐸 퐺 provides task- aware geometry risks, orientation guidance, parameter baselines, and design hints. The KG-based evidence 퐸 퐾퐺 provides ranked material candidates and supporting task- material-process evidence traces. By jointly using 푇 , 퐸 퐺 , and 퐸 퐾퐺 , the LLM performs constrained task-driven reasoning rather than unconstrained material or process recommendation. 3.3.1. KG-Based Material Evidence Generation The KG-based material evidence is generated by a task- conditioned KG scoring function: 퐸 퐾퐺 = 퐾퐺 (푇,푃,퐾퐺) = 푀 rank ,Π KG . (7) where 푃 is the fixed printer/process profile, 푀 rank denotes the ranked material candidates andΠ KG denotes the support- ing KG evidence traces. Algorithm 1 Task-conditioned KG material scoring Require: Task description 푇 , Task-Material-Process Knowledge Graph 퐾퐺, printer/process profile 푃 Ensure: KG-based material evidence package 퐸 퐾퐺 1: Activate task-relevant application and property nodes from 푇 . 2: Filter candidate materials using printer/process compat- ibility 푃 . 3: for each candidate material 푚 do 4: Retrieve KG evidence paths connecting activated ap- plication/property nodes to 푚. 5: Compute application and property evidence scores. 6: Compute material difficulty and printer/process feasi- bility penalties. 7: Compute the raw material score 푠 raw (푚). 8: Store material 푚, 푠 raw (푚), and supporting KG evi- dence traces. 9: end for 10: Rank candidate materials by 푠 raw (푚). 11: Return 퐸 퐾퐺 , including ranked material candidates and evidence traces. Algorithm 1 summarizes the procedure used by 퐾퐺 to produce 퐸 퐾퐺 . The material ranking in Algorithm 1 uses an additive evidence score: 푠 raw (푚) = 푠 app (푚)+푠 prop (푚)−푠 diff (푚)−푠 printer (푚). (8) The positive evidence terms are computed from task- activated application and property nodes: 푠 app (푚) = ∑ 푎∈(푇) 훼 푎 푤 app 푎,푚 , 푠 prop (푚) = ∑ 푝∈(푇) 훼 푝 푤 prop 푝,푚 . (9) Here, (푇) and (푇) denote the application and prop- erty nodes activated by the task description. The coefficients Du et al.: Preprint submitted to ElsevierPage 6 of 16 훼 denote activation strengths, and the weights 푤 denote heuristic evidence strengths linking activated nodes to mate- rial 푚. The penalty 푠 diff (푚) represents material printing diffi- culty, such as drying requirements, warping tendency, extru- sion stability, and tuning sensitivity. The penalty 푠 printer (푚) represents printer/process infeasibility and acts as a feasibil- ity constraint under the fixed experimental setup. Materials with positive raw scores are ranked and passed to the LLM together with their supporting KG evidence traces. Therefore, the LLM does not perform unconstrained material selection. Instead, material recommendation is grounded by task-conditioned KG evidence and further integrated with the geometry-grounded evidence package 퐸 퐺 during final report generation. 3.4. Post-Print Self-Improvement Module The post-print self-improvement module provides an op- tional mechanism for refining future recommendations using post-print outcomes. After printing, the user can provide feedback 퐹, such as print success or failure, visible defects, dimensional-fit issues, support-removal difficulty, material behavior, or an overall satisfaction rating. The framework stores this feedback together with the task intent, geometry- grounded evidence, generated report, selected material, and process settings: 퐻 = 퐺,푇,퐸 퐺 ,푂,퐹,푀,Θ,(10) where 푀 is the material used and Θ denotes the process parameters used for printing. The stored record 퐻 can be used to refine future recommendations at two levels. First, it provides case-level memory for later LLM reasoning, allowing the system to retrieve prior outcomes for similar tasks, materials, geometries, or failure modes. Second, the feedback record can support LLM-assisted refinement of both the geometry- grounded heuristics and the Task-Material-Process KG. On the geometry side, this module can adjust risk-trigger thresholds, orientation-scoring preferences, material- specific parameter baselines, and design-suggestion priorities. On the knowledge side, it can refine material- suitability scores, printing-difficulty penalties, task-property- material links, and retrieval preferences in the KG 4. Experimental Design 4.1. Benchmark 3D Models Conventional printability studies often use all-in-one benchmark artifacts that combine many geometric features, such as overhangs, bridges, thin walls, holes, and fine details, into a single model [7, 12]. While such artifacts are useful for testing manufacturing limits, they are less suitable for our task-driven setting because multiple failure modes are mixed together and the intended task is usually implicit. There- fore, it becomes difficult to evaluate whether an assessment changes because of task intent or because of geometry alone. Instead, we use a set of focused STL benchmark models, each isolating a representative printability challenge. As 1 cm Figure 4: Benchmark models for task-driven printability evalu- ation. shown in Figure 4 and Table 2, the benchmark set includes an overhang test, a dimensional accuracy test, a fine positive feature test, and a bridging test. These models allow us to pair specific geometric risks with task-specific descriptions and evaluate whether the framework generates appropriate risks, explanations, and material/parameter/design suggestions. Table 2 Benchmark 3D models. Benchmark model Primary challenge Overhang testUnsupported surfaces, support need, surface scarring, and orientation sensitivity. Dimensional accuracy test Dimensional fidelity, tolerance, and fit-related accuracy. Fine positive features test Small raised details, thin spikes, feature preservation, and local cool- ing sensitivity. Bridging testUnsupported spans, sagging risk, cooling sensitivity, and bridge- related settings. 4.2. Task Description Design Each focused STL benchmark is paired with two novice- style task descriptions, as shown in Table 3. The descriptions are written in non-technical language and describe desired task outcomes rather than explicit material properties or slicer settings. For each description, we assign an implicit task priority and a reference material as the ground-truth material label. The implicit priority indicates the main requirement that the method should infer from the user intent, while the reference material is used to evaluate material recommendation accu- racy. 4.3. Printer and Material Setup All experiments are conducted under a fixed printer profile using a Bambu Lab X1 Carbon (X1C). Therefore, the framework does not perform printer selection; instead, all material and parameter recommendations are conditioned on the available X1C setup. The material candidate set is limited to four laboratory- available materials: PLA, TPU 95A, PETG, and PC. As Du et al.: Preprint submitted to ElsevierPage 7 of 16 Table 3 Task descriptions, implicit task priorities, and reference material labels. Benchmark Model Task DescriptionImplicit PriorityReference Material Overhang test I want this as a small indoor desk display used at room temperature. It will not carry a load or be handled often. My main concern is a clean underside without obvious sagging or support marks. Surface appearance; clean underside PLA I want to leave this part inside a parked car, where it may become hot. It must remain rigid and keep the overhanging shape without bending or warping. Warm-environment use; shape retention PC Dimensional accuracy test This part will be inserted into and removed from a matching rigid part many times. The fit should remain firm and snug, and the edges should resist chipping or cracking. I do not want the part to feel soft or rubbery. Fit reliability; firm shape PETG This part must stretch or compress when pushed onto another part and return to its original shape after removal. It will be attached and removed repeatedly and must not split or remain deformed. Slight flexibility; crack avoidance TPU 95A Fine positive features test The small raised details will be touched and rubbed frequently and may occasionally catch on other objects. They should remain clear and firm without becoming soft, chipped, or cracked. Detail retention; firmness; chip/crack resistance PETG I want this to be part of a soft toy. The small raised parts should not feel hard, sharp, or uncomfortable when touched. Softness; safety; hand feel TPU 95A Bridging test I want this bridge-like part to look clean on the bottom. I do not want the bottom surface to droop or look rough. Clean underside; low sagging PLA I want this bridge-like part to stay on the dashboard inside the car and still keep its shape. It only needs to hold something light. Warm-environment use; shape retention PC Table 4 Operational task-suitability tests and binary pass criteria. Within each benchmark, the two rows follow the order of the corresponding task descriptions in Table 3. The same protocol was applied to all successfully fabricated samples from the compared methods. Benchmark Model Test ProcedurePass Criterion Overhang test After support removal, if supports are used, inspect the underside and measure the maximum sag relative to the intended underside profile. No detached strands or major support-contact damage, and maximum underside sag ≤ 0.5 m. Expose the sample to 85 ◦ C for 30 min, cool it at room temperature for 10 min, and measure the permanent tip displacement. No cracking or visible warping, and permanent tip displacement ≤ 2% of the overhang length. Dimensional accuracy test Perform 20 insertion and removal cycles using the same standardized mating part, followed by a 500 g retention test for 10 s. The mating part remains retained, with no visible loosening, chipping, or cracking. Perform 20 insertion and removal cycles using the same standardized mating part. Measure the critical dimension before testing and 60 s after the final cycle. No splitting or cracking, and dimensional recovery ≥ 95% of the initial value. Fine positive features test Move a 1.0-m-thick cardboard strip across the same raised-detail region for 20 cycles under a 300 g normal load. All target details remain distinguishable, with no visible chipping, cracking, detachment, or permanent deformation. Compress the raised features to approximately 50% of their initial height for 10 cycles using a flat-ended probe. Measure feature height 60 s after the final cycle. Height recovery ≥ 90%, with no cracking, permanent collapse, or sharp exposed edges. Bridging test Inspect the bridge underside and measure the maximum mid-span sag relative to the intended bridge profile. Mid-span sag ≤ 2% of the bridge span, with no detached strands or major surface discontinuities. Apply a 300 g center load while exposing the sample to 85 ◦ C for 30 min. Remove the load, cool for 10 min, and measure the residual mid-span deflection. No cracking, and residual mid-span deflection ≤ 2% of the bridge span. summarized in Table 5, these materials cover distinct task- relevant properties, including ease of printing, flexibility, toughness, and heat resistance. 4.4. Compared Methods / Baselines We compare four method settings. The rule-based base- line uses only the Geometry-Grounded Layer, without LLM reasoning or material-knowledge grounding. The pure LLM Du et al.: Preprint submitted to ElsevierPage 8 of 16 Table 5 Material properties Material Printability Flexibility Durability Heat resistance PLAHighLowMediumLow PETGMedium Medium MediumMedium TPU 95ALowHighHighMedium PCLowLowHighHigh baseline uses only the LLM with the task description and the raw ASCII STL input, without the explicit geometry- grounded evidence package or material selection knowledge graph. The LLM + KG baseline adds the structured Task- Material-Process Knowledge Graph to the pure LLM set- ting, but still does not use the Geometry-Grounded Layer. Our method combines the Geometry-Grounded Layer and the LLM Layer with structured material/printer knowledge, enabling the LLM to reason over both computed geometry evidence and task-relevant material constraints. The self- improvement module is evaluated separately because it re- quires post-print feedback or expert correction. For LLM-based settings, we evaluate three Gemini 2.5 backbones: Flash-Lite, Flash, and Pro. The four methods are tested under the same backbone settings, so that the comparison focuses on the effect of geometry and knowledge grounding rather than the choice of LLM backbone. 5. Evaluation Metrics Overview We evaluate the compared methods from four perspec- tives: printability evaluation, task suitability evaluation, ma- terial selection accuracy and expert evaluation of report quality. In addition, we assess the self-improvement module separately for the proposed framework. 5.1. Geometry-Level Printability Evaluation A trial is counted as successful if the generated rec- ommendation is executable and produces a physical print without a major geometry-level failure. Recommendations that omit essential information, such as material, orientation, or process parameters, are counted as printability failures. The observed Geometry-Level Printability rate is ̂푝 print = 푁 print 푁 all ,(11) where 푁 print is the number of successful prints and 푁 all is the total number of validation trials. The evaluation includes 96 trials from 8 task-conditioned scenarios, 4 methods, and 3 repetitions, giving푁 all = 24 per method. Of these trials, 75 proceeded to physical printing, while 21 non-executable recommendations were retained as printability failures. 5.2. Task Suitability Evaluation Task suitability evaluation is conducted on successfully printed samples from the physical printability validation. A sample is entered into task-level testing only if it is success- fully fabricated, because failed prints do not provide usable physical parts for downstream task suitability evaluation. Each successfully printed sample is evaluated using the task-suitability test defined in Table 4. These tests assess underside quality and sag, heat-induced shape retention, repeated mating and retention, dimensional recovery after repeated insertion and removal, raised-feature abrasion re- sistance, feature recovery and touch safety, bridge underside sag, and residual bridge deflection under combined heat and load. A sample is counted as task-suitable only if it satisfies the corresponding binary pass criterion in Table 4. We report two task-suitability metrics: the observed task- suitability rate among successful prints, ̂푝 (task∣print) , and the observed end-to-end task-suitable rate, ̂푝 e2e . They are com- puted as ̂푝 (task∣print) = 푁 task 푁 print ,(12) ̂푝 e2e = 푁 task 푁 all ,(13) where 푁 all , 푁 print , and 푁 task denote the total trials, suc- cessful prints, and task-suitable prints, respectively. In this study, 푁 all = 24 for each method. If 푁 print = 0, ̂푝 task∣print is reported as N/A. Confidence intervals. We report marginal Wilson 95% confidence intervals for ̂푝 print and ̂푝 e2e . Wilson intervals are used because the outcomes are binary and include boundary values near 0 or 1, where normal-approximation intervals may be unreliable. For ̂푝 = 푘∕푛, the interval is CI Wilson = ̂푝 + 푧 2 2푛 ± 푧 √ ̂푝(1−̂푝) 푛 + 푧 2 4푛 2 1 + 푧 2 푛 ,(14) where 푧 = 1.96 and 푛 = 24. These intervals describe the precision of the aggregate success-rate estimates under com- parable evaluation conditions, rather than instance-specific predictions. 5.3. Material Selection Accuracy Since material choice is coupled with both geometry- level printability and task suitability, material-selection behavior provides important evidence for interpreting the physical validation results. We therefore evaluate material- selection accuracy and analyze the corresponding confusion patterns to identify where each method succeeds or fails in task-conditioned material reasoning. Material selection accuracy measures whether the pri- mary recommended material matches the reference material label in Table 3. To assess prediction stability, each of the eight task descriptions was evaluated five times, yielding 40 material predictions for each LLM-based method under each backbone. With three LLM-based methods and three Gemini 2.5 backbones, this resulted in 360 LLM-based predictions. The deterministic rule-based baseline was evaluated over the Du et al.: Preprint submitted to ElsevierPage 9 of 16 Table 6 Expert evaluation rubric for structured reports. ComponentWhat it evaluatesScoring criterion Overall verdict 푦Whether the report gives a correct and task- aware printability judgment. 1: misleading or unsupported verdict; 5: accurate verdict that reflects geometry, task, and setup constraints. Material 푀Whether the recommended material is feasi- ble and appropriate for the task. 1: incompatible or task-irrelevant material; 5: cor- rect, feasible, and well-matched material choice. Risks 푅 푇 Whether the report identifies task-relevant printability and functional risks. 1: misses major risks or gives generic risks; 5: covers key risks and connects them to the task and geometry. Parameters ΘWhether the process guidance is feasible and useful for the given printer/material setup. 1: vague, infeasible, or unsafe parameters; 5: specific and practical parameters that address the identified risks. Design guidance 퐷 푇 Whether the report provides useful CAD or geometry modification suggestions. 1: no useful design guidance; 5: actionable design changes that directly improve printability or task suitability. Explanation 퐸Whether the report clearly explains the rea- soning behind the recommendation. 1: unclear or disconnected explanation; 5: clear reasoning linking task intent, geometry evidence, material choice, parameters, and risks. same 40 repeated cases for comparison, giving 400 material predictions in total. 5.4. Expert Evaluation of Report Quality Beyond geometry-level printability, task suitability, and material selection,the generated reports also explain why specific recommendations are made. Such explanations are important for novice users, but they are difficult to quantify with objective metrics. Therefore, we use expert evaluation to assess the quality of the generated reports. Expert evaluation assesses each structured report based on six output components(as Eq. (6)): overall verdict, ma- terial recommendation, task-relevant risks, parameter guid- ance, design guidance, and explanation. This evaluation fo- cuses on the quality and usefulness of the generated pre-print recommendations. It does not directly measure downstream functional performance, such as thermal deformation, snap- fit durability, surface softness, or long-term fit retention. The rubric of expert evaluation is shown in Table 6. Since expert scoring is time-consuming, we randomly select 10 cases from the 40 runs to reduce expert fatigue and potential scoring errors. For each selected case, three experts evaluate anonymized reports from the four compared meth- ods. The reports were anonymized and randomly shuffled prior to evaluation so that the experts were blinded to the generating methods. 5.5. Post-Print Self-Improvement Module Evaluation The post-print self-improvement module is evaluated only for the proposed framework, because the baselines do not include a post-print refinement mechanism. The eval- uation covers two types of problematic cases. We select 10 recommendation cases with material-selection errors or low expert scores to evaluate whether self-improvement improves material recommendation and report quality. We also apply the self-improvement module to the six failed physical print trials from the proposed framework to evaluate whether observed print failures can be corrected through regenerated recommendations. For each selected case, post-print outcome information is added to the framework, and the recommendation is regen- erated. The recorded information includes print outcomes, failure modes, material-related issues, and task-level suit- ability observations when available. The regenerated outputs are evaluated using the same criteria as the main evaluation, including material selection accuracy, printability, task suit- ability, and expert report quality. 6. Results and Analysis 6.1. Geometry-Level Printability Evaluation Table 7 shows that the proposed framework achieves the best overall printability performance among the compared methods. The most informative contrast is between LLM + KG and the proposed framework; although KG grounding improves material-related reasoning, it does not by itself pro- duce complete and physically executable recommendations. LLM + KG often fails to generate essential guidance such as orientation and support-related decisions, which leads to poor printability despite improved material reason- ing. This indicates that material knowledge alone is insuf- ficient for printability assistance when geometry-grounded fabrication evidence is missing. Pure LLM performs better than LLM + KG in the printability evaluation, but this does not indicate stronger grounded reasoning. A more likely explanation is that pure LLM can still produce generic fabrication suggestions from Du et al.: Preprint submitted to ElsevierPage 10 of 16 Table 7 Printability and task-suitability evaluation over physical valida- tion trials. Method ̂ 풑 print ̂ 풑 (task∣print) ̂ 풑 e2e Rule-based50.0% 66.7% 33.3% Pure LLM37.5% 55.6% 20.8% LLM + KG0.0%N/A0.0% Ours: LLM + KG + Geo 75.0% 88.9% 66.7% Rule-based Pure LLM LLM+KGOurs (a) Under-extrusion (b) Layer separation (c) Spaghettiing (d) Poor bed adhesion (f) Surface blobs(g) Bridging failure(h) Bridging failure(i) Under-extrusion Case 1 Case 2 Ours After Improve (e) No failure (j) No Failure Figure 5: Representative physical outcomes from geometry- level printability validation and post-print self-improvement. The first four columns show representative failure modes observed across the compared methods. The last column shows the corrected prints after applying the post-print self- improvement module to the failed cases. its internal prior knowledge, which helps maintain a rela- tively high complete-recommendation rate even when those suggestions are not well matched to the specific geometry or intended use. Its advantage is therefore one of generic completeness rather than geometry-grounded correctness. The proposed framework improves printability because it combines task-aware material reasoning with geometry- grounded evidence, including heuristic risk analysis and orientation guidance. This combination leads to more com- plete and more executable recommendations than those of the compared baselines. The remaining failures should be interpreted together with the material selection results, since some materials that are better aligned with the intended use are also more difficult to fabricate reliably. Figure 5 shows representative geometry-level failure cases observed during physical validation, illustrating the types of fabrication failures counted in the printability evaluation. 6.2. Task Suitability Evaluation Table 7 also reports the task-suitability results. Our method achieves the highest task-suitability rate among suc- cessfully printed samples, with ̂푝 (task∣print) = 88.9% (16∕18). It also achieves the highest end-to-end task-suitable rate, with ̂푝 e2e = 66.7% (16∕24), compared with 33.3% for the rule-based baseline, 20.8% for Pure LLM, and 0.0% for LLM + KG. These results indicate that the proposed framework improves both physical printability and the likelihood that a fabricated part satisfies its intended task requirement. Figure 6(a) shows that our method occupies the upper- right region and is the only non-dominated method based on the observed rates. The dashed arrows indicate the direction of stronger joint performance. The Wilson intervals describe the uncertainty in the underlying rates for comparable trials, rather than the outcome of an individual future trial. Figure 6(b) reveals different benchmark-level patterns. Rule-based remains at 50.0% printability across all four benchmark CAD models, while Pure LLM is more benchmark-dependent, with printability distributed across 0.0%, 50.0%, and 100.0%. All LLM + KG points remain at 0.0%. In contrast, our method maintains at least 50.0% printability on every benchmark CAD model, and three of the four models achieve an end-to-end task-suitable rate of at least 50.0%. The strongest results are obtained for bridging, with 100% on both objectives, and fine positive features, with 100% printability and 83.3% end-to-end task suitability. 6.3. Material Selection Accuracy Figure 7 summarizes the material selection accuracy across methods and LLM backbones. The proposed frame- work achieves the best overall result, with the highest ac- curacy obtained by Gemini 2.5 Flash-Lite, where accuracy improves from 37.5% under pure LLM prompting to 90.0%. The confusion matrices also show more stable predictions, with fewer off-category outputs and fewer missing recom- mendations. The progression across methods is informative. Under Gemini 2.5 Flash-Lite, pure LLM prompting performs sim- ilarly to the rule-based baseline, suggesting that free-form LLM reasoning alone does not provide sufficient grounded information for reliable material prediction. Adding the KG substantially improves accuracy, indicating that structured material knowledge is a major source of evidence for this task. The proposed framework further improves upon pure LLM + KG, suggesting that geometry-grounded rule-based evidence and heuristic priors provide additional complemen- tary information. Together, these results support the frame- work design, in which KG grounding strengthens material reasoning and geometry-grounded evidence further refines task-conditioned material selection. The remaining errors of the proposed framework are mainly concentrated in PLA and PETG confusion, rather than larger mismatches involving TPU 95A or PC. This pattern is understandable because PLA and PETG are rel- atively similar within the evaluated material set and can both satisfy some rigid-part requirements, although PETG is typically tougher while PLA is easier to print. This suggests residual ambiguity between nearby material choices rather than arbitrary failure, while materially distinct categories remain more reliably separated. A further trend can also be observed across the evaluated LLM backbones. In the pure LLM and pure LLM + KG settings, Gemini Flash generally performs better than Flash- Lite, which may indicate that the larger model contains Du et al.: Preprint submitted to ElsevierPage 11 of 16 0%25%50%75%100% f 1 : Geometry-Level Printability 0% 25% 50% 75% 100% f 2 : End-to-End Task-Suitable Rate better (a) Rule-based Pure LLM LLM + KG Ours: LLM + KG + Geo 0%25%50%75%100% f 1 : Geometry-Level Printability (b) Method (color) Rule-based Pure LLM LLM + KG Ours: LLM + KG + Geo Benchmark CAD (marker) Overhang Dimensional accuracy Fine positive features Bridging Figure 6: Joint comparison of geometry-level printability and task suitability under Gemini 2.5 Flash-Lite. The horizontal axis reports geometry-level printability, 푓 1 = ̂푝 print , and the vertical axis reports the end-to-end task-suitable rate, 푓 2 = ̂푝 e2e . In (a), each cross aggregates 24 trials for one method, with horizontal and vertical error bars showing the corresponding Wilson 95% confidence intervals. These intervals quantify uncertainty in the estimated success rates and indicate the range of underlying performance compatible with the observations for future comparable trials. In (b), each point aggregates six trials for one method-benchmark CAD combination. Colors distinguish the four methods, while marker shapes distinguish the four benchmark CAD models. Exactly overlapping points are slightly offset for visibility. Higher values on both axes indicate better joint performance. richer internal knowledge relevant to 3D printing and ma- terial selection. However, this trend does not continue from Flash to Pro. Instead, performance decreases at the largest scale, which may reflect an inverse-scaling [27, 38, 28] effect in the present evaluation setting. A possible explanation is that larger models are more likely to generate alternative but still plausible material choices based on broader prior knowledge, even when the evaluation uses a fixed reference material label [39]. 6.4. Expert Evaluation of Report Quality Figure 8 shows the expert evaluation results across the six structured report components. The proposed framework achieves the highest mean score in all components, indicat- ing that the generated reports are more complete, task-aware, and actionable than those produced by the three baselines. The rule-based baseline receives lower scores because it mainly relies on local geometry heuristics. It can provide basic printability warnings, but it has limited ability to infer the user’s functional intent, select task-appropriate materials, or explain trade-offs among material, geometry, and process settings. The pure LLM baseline improves over the rule- based baseline in most components, especially in verdict generation and explanation. However, its performance is less stable in material selection and design guidance, suggesting that free-form LLM reasoning may produce plausible but insufficiently grounded recommendations. The proposed framework performs best in material recommendation, task-relevant risk identification, parameter guidance, and explanation quality. These improvements show the benefit of combining geometry-grounded evidence with structured material and printer knowledge. In particular, the higher scores for risks and parameters suggest that the framework can better connect task intent with geometry- dependent failure modes and feasible slicer-level actions. In addition to the aggregate expert scores, Figure 9 shows a representative same-geometry different-task case. The same dimensional-accuracy STL leads to different rec- ommendations under rigid and flexible fitting intents: PETG and higher-stiffness settings for the rigid fitting case, but TPU 95A and more flexible tuning for the snap-fit case. This example qualitatively illustrates the task-conditioned behavior measured by the evaluation. 6.5. Post-Print Self-Improvement Module Evaluation Figure 10(a) shows that material recommendation accu- racy on the 10 selected problematic recommendation cases improves from 0/10 before self-improvement to 7/10 after self-improvement. Figure 10(b) shows that expert scores improve across all six structured report components, indi- cating that self-improvement improves the coherence and usefulness of the regenerated reports. Du et al.: Preprint submitted to ElsevierPage 12 of 16 Figure 7: Material selection accuracy and confusion matrices across evaluation methods. The bar chart reports material recommendation accuracy under different methods across three Gemini model scales. Error bars indicate standard deviation across five runs. The confusion matrices show primary material predictions for each LLM-based method and model scale, where rows denote ground-truth materials and columns denote predicted primary materials. Asterisks mark correct predictions. Figure 8: Expert evaluation of structured printability reports. Du et al.: Preprint submitted to ElsevierPage 13 of 16 Inputs Geometry:Task: “This part will be inserted into and removed from a matching rigid part many times. The fit should remain firm and snug, and the edges should resist chipping or cracking. I do not want the part to feel soft or rubbery.” OverallLikely printable with tuning. Inferred priorityFit reliability; firm shape. MaterialPETG Why this material PETG offers an excellent balance of stiffness, toughness, and dimensional stability, which is crucial for parts that need to fit firmly and resist loosening over repeated use. Key process parameters • Orientation: Rotate so Y points up • Nozzle: 0.4 m • Layer height: 0.16 m • Walls: 4 walls • Sparse infill density: 20% • Sparse infill pattern: Gyroid • Support: Treesupport, build plate only • Bed adhesion: Brim (4–6 m) • Print speed: Outer wall 35–45 m/s; inner wall 60–80 m/s • Cooling: Part cooling fan 30–50% Design suggestions • Replace unsupported undersides with chamfers, arches, or self-supporting angles • Add a small flat landing pad or sacrificial print foot • Add a dedicated print face so the intended orientation is obvious and stable Risks • High —Dimensional accuracy for interlocking fit: print a small calibration piece first and design slight clearances (0.1–0.2 m). • High —Poor bed adhesion in optimized orientation: use a brim or raft; consider a sacrificial print foot. • Medium —Part loosening over repeated use: prioritize PETG or PC; avoid PLA; use 4 walls and 20% infill. • Medium —Overhang regions may need support: enable tree supports and inspect support placement. • Medium —Suboptimal initial orientation: rotate the model so Y points up. Case A: Rigid fitting part Inputs Geometry:Task: “This part must stretch or compress when pushed onto another part and return to its original shape after removal. It will be attached and removed repeatedly and must not split or remain deformed.” Overall Highly printable with appropriate material and orientation choices Inferred prioritySlight flexibility; crack avoidance MaterialTPU 95A Why this material TPU 95A offers excellent flexibility and high impact resistance, directly addressing the need for a part that will not crack if pressed a little too hard and can snap or push onto another part. Key process parameters • Orientation: Keep current orientation (Z+) • Nozzle: 0.4 m • Layer height: 0.16 m • Walls: 3 walls • Sparse infill density: 15–20% • Sparse infill pattern: Gyroid • Support: Tree support, build plate only • Bed adhesion: No brim or a small 3 m brim • Print speed: Outer wall 25–45 m/s, inner wall 60–80 m/s • Cooling: Default for TPU (typically 50–70% fan speed) Design suggestions • Refine snap-fit tolerances for FDM and the chosen material • Add chamfers or fillets to sharp overhangs and bottom edges Risks • High — Dimensional accuracy for snap-fit features with flexible materials: print a small test piece first and tune tolerances. • Medium — Overhang regions will require support structures: use tree support, build plate only. • Low — TPU printing can be slower and require more tuning: start with the default TPU 95A HF profile and ensure the filament is dry. Case B: Flexible fitting part Figure 9: Same geometry under different task case study. The same STL geometry is evaluated under two novice-style task descriptions with different functional requirements. For the rigid fitting task, the framework infers fit reliability and firm shape as the main priorities and recommends PETG with higher wall count, moderate infill, brim-assisted adhesion, and a Y-up candidate orientation to support stiffness and repeated-use stability. For the flexible fitting task, the framework instead prioritizes slight flexibility and crack avoidance, recommending TPU 95A with the current Z+ orientation, lower infill, and TPU-specific tuning. This comparison shows that the proposed framework adapts material, parameter, design, and risk recommendations according to task intent rather than geometry alone. For physical validation, the self-improvement module is applied to the six failed print trials from the proposed frame- work. These failures correspond to two task-conditioned benchmark scenarios with three repeated trials each. After self-improvement, the failed trials are corrected, increasing the proposed framework’s printability from 75.0% to 100.0%. Since the corrected printed samples also pass the correspond- ing task-suitability tests, the end-to-end task-suitable rate ̂푝 e2e also improves from 66.7% to 91.7%. Figure 5 shows two representative physical outcomes before and after self-improvement. These examples qualita- tively illustrate how observed post-print failures can be used to improve subsequent recommendations. 7. Discussion The results show that task-driven printability assistance requires both geometry-grounded fabrication evidence and task-material-process knowledge. Geometry evidence makes the recommendation physically executable, while (a)(b) Figure 10: Effect of self-improvement module. structured material/printer knowledge helps align the recommendation with the intended use. The weak printability of the LLM + KG baseline further indicates that material knowledge alone is insufficient without orientation, support, adhesion, and other geometry-dependent evidence. The remaining failures reflect the trade-off between task suitability and fabrication difficulty. Materials that better satisfy functional requirements, such as flexibility or heat Du et al.: Preprint submitted to ElsevierPage 14 of 16 resistance, can also introduce higher printing risks. The current evaluation is limited to four focused STL models and four laboratory-available materials, so broader validation is needed across more geometries, printers, materials, and downstream functional tests. For a single Gemini 2.5 Flash-Lite request, the geometry- grounded layer takes 0.003 s, KG-based material reasoning takes 5.350 s, and LLM inference takes 40.518 s, resulting in an end-to-end runtime of 45.871 s. The main bottleneck is LLM inference, while local geometry analysis adds negli- gible overhead. 8. Conclusion This paper presented a task-driven 3D printability assis- tance framework that combines geometry-grounded analy- sis, structured material/printer knowledge, and constrained LLM reasoning. Experiments show improved material se- lection, physical printability, task suitability, and expert- rated report quality compared with rule-based, pure LLM, and LLM + KG baselines. The post-print self-improvement evaluation further suggests that observed failures can refine subsequent recommendations, demonstrating the feasibility of geometry- and knowledge-grounded LLM reasoning for interactive, task-aware printability assistance. Declaration of Generative AI Use ChatGPT (OpenAI) was used solely to improve the lan- guage and readability of this manuscript. The authors re- viewed all revisions and take full responsibility for the content. Online Supplementary Material Representative interactive outputs and case demonstra- tions are available on the project page. References [1] Ahmad, N.N., Wong, Y.H., Ghazali, N.N.N., 2022. A systematic review of fused deposition modeling process parameters. Soft Science 2, 11. [2] Altıparmak, S.C., Daminabo, S.I.C., 2024. Suitability analysis for extrusion-based additive manufacturing process. Additive Manufac- turing Frontiers 3, 200106. [3] Bian, H., 2025. LLM-empowered knowledge graph construction: A survey. arXiv preprint arXiv:2510.20345 . [4] Booth, J.W., Alperovich, J., Chawla, P., Ma, J., Reid, T.N., Ramani, K., 2017. The design for additive manufacturing worksheet. Journal of Mechanical Design 139, 100904. [5] Chandrasekhar, A., Chan, J., Ogoke, F., Ajenifujah, O., Farimani, A.B., 2024. AMGPT: A large language model for contextual query- ing in additive manufacturing. Additive Manufacturing Letters 11, 100232. [6] Chen, Y., Wang, H., Wu, Y., Wang, H., 2020. Predicting the print- ability in selective laser melting with a supervised machine learning method. Materials 13, 5063. [7] Decker, N., Yee, A., 2015. A simplified benchmarking model for the assessment of dimensional accuracy in FDM processes. International Journal of Rapid Manufacturing 5, 145–154. [8] Du, Z., Bowman, M., Zheng, Q., Zhang, X., 2026. Decision- aware uncertainty evaluation of vision-language model-based early action anticipation for human-robot interaction. arXiv preprint arXiv:2603.10061 . [9] Eslaminia, A., Jackson, A., Tian, B., Stern, A., Gordon, H., Malhotra, R., Nahrstedt, K., Shao, C., 2025. FDM-Bench: a domain-specific benchmark for evaluating large language models in additive manufac- turing. Manufacturing Letters 44, 1415–1424. [10] Frank, D., Fadel, G., 1995. Expert system-based selection of the preferred direction of build for rapid prototyping processes. Journal of Intelligent Manufacturing 6, 339–345. [11] Fu, Y., Downey, A., Yuan, L., Pratt, A., Balogun, Y., 2021. In situ monitoring for fused filament fabrication process: A review. Additive Manufacturing 38, 101749. [12] Fudos, I., Ntousia, M., Stamati, V., Charalampous, P., Kontodina, T., Kostavelis, I., Tzovaras, D., Bilalis, L., 2021. A characterization of 3D printability. Computer-Aided Design and Applications 18, 1279– 1295. [13] Goala, S., Sarkar, P., 2024. Selection methodology for additive manufacturing processes considering design, material, and manufac- turability. Rapid Prototyping Journal 30, 15–32. [14] Gobert, C., Reutzel, E.W., Petrich, J., Nassar, A.R., Phoha, S., 2018. Application of supervised machine learning for defect detection dur- ing metallic powder bed fusion additive manufacturing using high resolution imaging. Additive Manufacturing 21, 517–528. [15] Guo, Q., Ye, J., Lu, H., Quan, G., Wang, Z., Zhao, Y., Xie, Y.M., 2024. Design and validation of 3D self-supporting structures and printing paths for multi-axis additive manufacturing. Additive Manufacturing 96, 104563. [16] Hammoud, A., Mohomad, Y., Shomar, H., Masad, E., Arroyave, R., Tafreshi, R., 2026. Data-driven framework for printability and geometric quality prediction in 3D concrete printing. Materials & Design 261, 115309. [17] Han, X., Qin, Q., Chen, S., Liu, X., Cui, L., 2025. Support-free 3D printing based on model decomposition. Micromachines 16, 1316. [18] Henn, J., Hauptmannl, A., Gardi, H.A., 2025. Evaluating the print- ability of STL files with ML. arXiv preprint arXiv:2509.12392 . [19] Jadhav, Y., Pak, P., Farimani, A.B., 2025. LLM-3D print: Large language models to monitor and control 3D printing. Additive Manufacturing 114, 105027. [20] Jiang, J., Xu, X., Stringer, J., 2018. Support structures for additive manufacturing: A review. Journal of Manufacturing and Materials Processing 2, 64. [21] Jiang, J., Xu, X., Stringer, J., 2019. Optimization of process planning for reducing material waste in extrusion based additive manufacturing. Robotics and Computer-Integrated Manufacturing 59, 317–325. [22] Křivohlavý, P., Škaroupka, D., Krejčiřík, P., Krčma, M., Koutný, D., 2025. Influence of printhead incline on the accuracy of robotic 3D printed spatial polymer struts. The International Journal of Advanced Manufacturing Technology 141, 579–589. [23] Li, X., Liu, W., Hu, Z., He, C., Ding, J., Chen, W., Wang, S., Dong, W., 2024. Supportless 3D-printing of non-planar thin-walled structures with the multi-axis screw-extrusion additive manufacturing system. Materials & Design 240, 112860. [24] Li, Y., Zhao, H., Jiang, H., Pan, Y., Liu, Z., Wu, Z., Shu, P., Tian, J., Yang, T., Xu, S., et al., 2026. Large language models for manufactur- ing. Journal of Manufacturing Systems 86, 516–545. [25] Lu, T., 2016. Towards a fully automated 3D printability checker, in: 2016 IEEE International Conference on Industrial Technology (ICIT), IEEE. p. 922–927. [26] Luo, L., Baran, I., Rusinkiewicz, S., Matusik, W., 2012. Chopper: Partitioning models into 3D-printable parts. ACM Transactions on Graphics (TOG) 31, 1–9. [27] McKenzie, I.R., Lyzhov, A., Pieler, M., Parrish, A., Mueller, A., Prabhu, A., McLean, E., Kirtland, A., Ross, A., Liu, A., et al., 2023. Inverse scaling: When bigger isn’t better. arXiv preprint arXiv:2306.09479 . Du et al.: Preprint submitted to ElsevierPage 15 of 16 [28] Michaelov, J., Bergen, B., 2023. Rarely a problem? language models exhibit inverse scaling in their predictions following few-type quanti- fiers, in: Findings of the Association for Computational Linguistics: ACL 2023, p. 14162–14174. [29] Mycroft, W., Katzman, M., Tammas-Williams, S., Hernandez-Nava, E., Panoutsos, G., Todd, I., Kadirkamanathan, V., 2020. A data-driven approach for predicting printability in metal additive manufacturing processes. Journal of Intelligent Manufacturing 31, 1769–1781. [30] Ouerghemmi, C., Ertz, M., 2025. Integrating large language models into digital manufacturing: A systematic review and research agenda. Computers 14, 318. [31] Pak, P., Farimani, A.B., 2025. AdditiveLLM: Large language models predict defects in metals additive manufacturing. Additive Manufac- turing Letters 14, 100292. [32] Pradel, P., Zhu, Z., Bibb, R., Moultrie, J., 2018. Investigation of design for additive manufacturing in professional design practice. Journal of Engineering Design 29, 165–200. [33] Sahar, T., Rauf, M., Murtaza, A., Khan, L.A., Ayub, H., Jameel, S.M., Ahad, I.U., 2023. Anomaly detection in laser powder bed fusion using machine learning: A review. Results in Engineering 17, 100803. [34] Taufik, M., Jain, P.K., 2013. Role of build orientation in layered manufacturing: a review. International Journal of Manufacturing Technology and Management 27, 47–73. [35] Trovato, M., Cicconi, P., 2024. A decision tree approach for an early evaluation of 3D models in design for additive manufacturing. Procedia CIRP 128, 96–101. [36] Uz Zaman, U.K., Rivette, M., Siadat, A., Mousavi, S.M., 2018. Inte- grated product-process design: Material and manufacturing process selection for additive manufacturing using multi-criteria decision making. Robotics and Computer-Integrated Manufacturing 51, 169– 180. [37] Vanek, J., Galicia, J.A.G., Benes, B., 2014. Clever support: Efficient support structure generation for digital fabrication. Computer Graph- ics Forum 33, 117–125. [38] Wei, J., Kim, N., Tay, Y., Le, Q., 2023. Inverse scaling can become u-shaped, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 15580–15591. [39] Yang, M., Huang, E., Zhang, L., Surdeanu, M., Wang, W.Y., Pan, L., 2025. How is LLM reasoning distracted by irrelevant context? An analysis using a controlled benchmark, in: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 13340–13358. [40] Zheng, Q., Zhang, J., Gockel, J., Wakin, M.B., Brice, C., Zhang, X., 2026a. QA-VLM: Providing human-interpretable quality assessment for wire-feed laser additive manufacturing parts with vision language models. Journal of Manufacturing Processes 160, 611–623. [41] Zheng, Q., Zhang, J., Stebner, A., Zhang, X., 2026b. In-context- learning-assisted vision-language models for quality assessment in metal additive manufacturing. JOM Online first. Du et al.: Preprint submitted to ElsevierPage 16 of 16