Task-Driven 3D Printability Assistance via Geometry- and Knowledge-Grounded LLM Reasoning

Introduction

• Existing automated methods identify geometry risks but do not infer intended use or jointly recommend materials and processes; prior work recognizes printability and task suitability as distinct objectives [12].A geometry checker may deem a model printable with PLA even when heat exposure in the intended task could soften or deform the part.Source: §1 · p. 1
• Task suitability has often relied on expert material and process selection followed by post-print evaluation and iteration, making success expertise-dependent and requiring repeated prints.Source: §1 · p. 1
• The framework combines structured geometry evidence, a Task-Material-Process knowledge graph, and constrained LLM reasoning to recommend printability, risks, materials, parameters, design guidance, and explanations.An optional post-print module records outcomes and ratings to update memory, heuristics, and future scoring.Source: §1 · p. 2
• The task-driven formulation allows the same geometry to receive different material, parameter, design, and risk recommendations depending on the user’s intended task.Source: §1 · p. 2
• The evaluation spans multiple LLM backbones, novice-style tasks, same-geometry different-task cases, printability, task suitability, material accuracy, report quality, and selected self-improvement cases.Source: §1 · p. 2

2.1. Rule-Based Printability Assessment

• Rule-based checkers evaluate CAD or STL geometry against printer constraints, while printability also depends on orientation, support strategy, and process planning [12].These methods provide interpretable and reproducible manufacturability evidence, but typically assume the printer, material, orientation objective, and process constraints in advance; they generally do not infer task intent or adapt recommendations to task suitability.Source: §2.1 · pp. 2–3

2.2. Data-Driven Printability Prediction

• Data-driven methods learn printability from geometry and process conditions, but their training, explanation, and transfer limitations constrain task-driven assistance [18].Approaches include decision trees and neural networks using CAD features, STL representations, or printed artifacts; process-aware work also predicts geometric quality, selects parameters, and models process windows or in-situ defects. These models may require substantial training data, offer limited explanations, and be tied to particular machines, materials, process windows, geometry families, or quality metrics.Source: §2.2 · pp. 3–4

2.3. LLM-Based Manufacturing Assistance

• LLMs can interpret natural-language instructions and explain manufacturing decisions, but existing systems focus on querying, monitoring, or defect diagnosis rather than pre-print assistance conditioned on an STL and task [8] [19] [5].Examples include AMGPT for contextual querying [5], FDM-Bench for evaluating domain-specific competence, AdditiveLLM for defect prediction, and LLM-3D Print for print-image monitoring and corrective actions [19].Source: §2.3 · p. 4

3. Methodology

• The three-layer framework combines geometry-grounded evidence, a task-material-process knowledge graph, and LLM reasoning, with an optional module for post-print refinement from outcomes and user feedback.The integrated reasoning targets recommendations that are both geometrically printable and aligned with the intended task.Source: §3 · p. 4

3.1. Geometry-Grounded Layer

• The Geometry-Grounded Layer converts an STL and lightweight task cues into a structured evidence package through mesh computation, orientation scoring, and heuristic evidence generation.This package is passed to downstream LLM reasoning.Source: §3.1 · p. 4

3.1.1. Mesh-level Computation

• Mesh analysis parses the STL as a triangle mesh and computes dimensions, surface area, estimated volume, watertightness, boundary and non-manifold edges, and degenerate triangles.These checks provide evidence about geometry validity and slicing suitability [12].Source: §3.1.1 · p. 4

3.1.2. Orientation Scoring

• Orientation scoring evaluates six canonical build orientations using support-risk, base-contact, slenderness, and height-dominance indicators, then selects the orientation with the lowest score.The score weights were empirically set in preliminary validation experiments on primitive geometries to balance the four indicators.Source: §3.1.2 · pp. 4–5
• The selected orientation and its score join the geometry-grounded evidence package for downstream heuristic generation and LLM reasoning.The method passes both the orientation-dependent score and selected candidate onward.Source: §3.1.2 · p. 5

3.1.3. Heuristic Evidence and Scores

• Rule-triggered heuristics convert geometry evidence and task cues into risk warnings, CAD design hints, initial slicer-parameter baselines, and keyword-based material candidates.Examples include support warnings for high support risk, adhesion guidance for weak base contact, mesh-repair warnings for invalid topology, and stability warnings for high slenderness; task cues reprioritize risks such as appearance, fit, or heat concerns.Source: §3.1.3 · p. 5

3.2. Task-Material-Process Knowledge Graph

• The Task-Material-Process KG contains 52 typed nodes and 347 typed edges linking applications, material properties, candidate materials, and printer/process constraints.Its node counts are 8 application, 12 material-property, 7 printer/process constraint, and 25 candidate-material nodes; its edge counts include 72 application-material, 96 property-material, and 179 printer/material feasibility edges.Source: §3.2 · p. 5
• Experts defined a fixed KG schema, while the LLM organized official manufacturer documentation [3] and suggested links and evidence strengths that experts normalized and checked.The resulting graph is a schema-constrained material-printer knowledge layer, not an unconstrained LLM-generated knowledge base.Source: §3.2 · p. 5

3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer

• The reasoning layer combines the task description, geometry-grounded evidence, and KG-based material evidence to generate a structured report with a verdict, material recommendation, risks, process guidance, design guidance, and rationale.Each report component can be traced to task intent, geometry evidence, or graph-retrieved material/process knowledge.Source: §3.3 · p. 6

3.3.1. KG-Based Material Evidence Generation

• KG scoring activates task-relevant application and property nodes, filters candidates for printer/process compatibility, retrieves evidence paths, and ranks materials with supporting traces.The procedure stores each candidate's raw score and evidence traces in the returned material evidence package.Source: §3.3.1 · p. 6
• The raw material score combines task-linked application and property evidence with penalties for material printing difficulty and printer/process infeasibility.Difficulty examples include drying requirements, warping tendency, extrusion stability, and tuning sensitivity; infeasibility acts as a constraint under the fixed experimental setup.Source: §3.3.1 · pp. 6–7
• Positive-score candidates and KG evidence traces ground material selection, which the LLM integrates with geometry evidence rather than recommending materials or processes unconstrained.Geometry evidence supplies task-aware risks, orientation guidance, parameter baselines, and design hints; KG evidence supplies ranked candidates and task-material-process traces.Source: §3.3 · p. 6, §3.3.1 · p. 7

3.4. Post-Print Self-Improvement Module

• Users can provide post-print feedback on outcomes such as success, defects, fit, support removal, material behavior, or overall satisfaction.The module is optional and collects feedback after printing.Source: §3.4 · p. 7
• The system stores feedback alongside task intent, geometry evidence, the generated report, selected material, and printing process settings.The stored record identifies the material used and the process parameters used for printing.Source: §3.4 · p. 7
• Stored records support case-level retrieval for similar tasks, materials, geometries, and failure modes, and can refine geometry heuristics and the Task-Material-Process knowledge graph.Refinements can adjust risk thresholds, orientation preferences, material-specific parameter baselines, design-suggestion priorities, suitability scores, difficulty penalties, task-property-material links, and retrieval preferences.Source: §3.4 · p. 7

4.1. Benchmark 3D Models

• Focused STL models isolate overhang, dimensional-accuracy, fine-feature, and bridging challenges for evaluating task-specific risks and recommendations.The design avoids combining multiple failure modes in one artifact, allowing task intent and geometry-specific effects to be assessed separately.Source: §4.1 · p. 7

4.2. Task Description Design

• Each benchmark has two non-technical task descriptions, with an implicit task priority and reference material assigned to each description.Descriptions specify desired outcomes rather than explicit material properties or slicer settings; reference materials serve as ground-truth labels for recommendation accuracy.Source: §4.2 · p. 7

4.3. Printer and Material Setup

• Experiments use a fixed Bambu Lab X1 Carbon profile and four candidate materials: PLA, TPU 95A, PETG, and PC.Printer selection is outside the evaluation; material and parameter recommendations are conditioned on the X1C setup.Source: §4.3 · p. 7

4.4. Compared Methods / Baselines

• Four settings isolate grounding effects: geometry-only rules, pure LLM, LLM+KG, and the proposed geometry-plus-knowledge method; self-improvement is evaluated separately.LLM-based settings use Gemini 2.5 Flash-Lite, Flash, or Pro backbones, held constant across methods; self-improvement requires post-print feedback or expert correction.Source: §4.4 · pp. 8–9

5.1. Geometry-Level Printability Evaluation

• Success required an executable recommendation producing a physical print without major geometry failure; omissions of essential inputs counted as failures across 96 trials, of which 75 reached printing and 21 were non-executable.The trials covered 8 task-conditioned scenarios, 4 methods, and 3 repetitions, for 24 trials per method.Source: §5.1 · p. 9

5.2. Task Suitability Evaluation

• Task suitability is tested only on successfully fabricated samples against Table 4 binary criteria, with conditional and end-to-end rates reported; marginal Wilson 95% intervals are provided for printability and end-to-end rates.The tests cover physical properties such as sag, heat-induced shape retention, repeated mating, dimensional recovery, abrasion resistance, and touch safety.Source: §5.2 · p. 9

5.3. Material Selection Accuracy

• Material-selection accuracy measures whether the primary recommendation matches the reference label; 40 predictions per LLM-based method-backbone pairing and 400 predictions overall support accuracy and confusion analyses.Eight task descriptions were each evaluated five times; the total includes 360 LLM-based predictions and 40 repeated rule-based cases.Source: §5.3 · pp. 9–10

5.4. Expert Evaluation of Report Quality

• Three blinded experts score reports on six components, using anonymized, shuffled reports from four methods across 10 randomly selected cases; this assesses recommendation quality, not downstream functional performance.The components are verdict, material recommendation, task-relevant risks, parameter guidance, design guidance, and explanation.Source: §5.4 · p. 10

5.5. Post-Print Self-Improvement Module Evaluation

• Post-print self-improvement is evaluated only for the proposed framework on 10 problematic recommendations and six failed physical trials, regenerating recommendations with recorded outcomes and assessing them by the main evaluation criteria.Problematic recommendation cases involve material-selection errors or low expert scores; added feedback includes failure modes, material issues, and task-suitability observations when available.Source: §5.5 · p. 10

6.1. Geometry-Level Printability Evaluation

• The proposed framework achieves the best overall printability among compared methods by combining task-aware material reasoning with geometry-grounded risk analysis and orientation guidance.The remaining failures should be interpreted alongside material selection: materials better aligned with intended use may be harder to fabricate reliably.Source: §6.1 · pp. 10–11, Table 7 · p. 11
• Pure LLM outperforms LLM + KG in printability through generic completeness, while LLM + KG often lacks orientation and support guidance for geometry-grounded, executable recommendations.This contrast indicates that material knowledge alone is insufficient; generic suggestions may not match the specific geometry or intended use.Source: §6.1 · pp. 10–11

6.2. Task Suitability Evaluation

• 88.9% (16/18) is the proposed method’s p̂(task∣print), and 66.7% (16/24) is its p̂e2e, versus 33.3% for rule-based, 20.8% for Pure LLM, and 0.0% for LLM + KG.The framework also improves physical printability and the likelihood that a fabricated part satisfies its intended task.Source: §6.2 · p. 11, Figure 6 · p. 12
• The proposed method maintains at least 50.0% printability on every benchmark CAD model; bridging reaches 100% on both objectives, while fine positive features reach 100% printability and 83.3% end-to-end suitability.Rule-based printability remains 50.0% across all four models, Pure LLM ranges from 0.0% to 100.0%, and LLM + KG remains at 0.0%.Source: §6.2 · p. 11, Figure 6 · p. 12

6.3. Material Selection Accuracy

• Gemini 2.5 Flash-Lite material-selection accuracy rises from 37.5% under pure LLM prompting to 90.0% with the proposed framework, the highest reported accuracy.KG grounding improves on pure LLM, and geometry-grounded evidence and heuristic priors further improve task-conditioned selection.Source: §6.3 · p. 11, Figure 7 · p. 13
• Proposed-framework material errors mainly confuse PLA with PETG; Gemini Flash generally exceeds Flash-Lite in pure LLM and LLM + KG settings, but performance decreases from Flash to Pro.PLA and PETG can both meet some rigid-part requirements, though PETG is typically tougher and PLA easier to print.Source: §6.3 · pp. 11–12

6.4. Expert Evaluation of Report Quality

• The proposed framework earns the highest mean expert score across all six report components, with strengths in material recommendations, task-relevant risks, parameter guidance, and explanations.For the same geometry, rigid-fitting intent yields PETG and higher-stiffness settings, while flexible-fitting intent yields TPU 95A and more flexible tuning.Source: §6.4 · p. 12, Figure 8 · p. 13, Figure 9 · p. 14

6.5. Post-Print Self-Improvement Module Evaluation

• On the 10 selected problematic recommendation cases, material recommendation accuracy improves from 0/10 before self-improvement to 7/10 after self-improvement.Expert scores improve across all six structured report components, indicating improved coherence and usefulness of regenerated reports.Source: §6.5 · p. 12

• Task-driven printability assistance requires both geometry-grounded fabrication evidence and task–material–process knowledge to make recommendations physically executable.Source: §7 · p. 14
• Functional materials offering flexibility or heat resistance can raise printing risks, and evaluation across four focused STL models and four available materials needs broader validation.Source: §7 · pp. 14–15
• 45.871 s end-to-end runtime comprises 0.003 s for geometry analysis, 5.350 s for KG-based material reasoning, and 40.518 s for LLM inference.LLM inference is the main bottleneck, while local geometry analysis adds negligible overhead.Source: §7 · p. 15

8. Conclusion

• The framework combines geometry-grounded analysis, structured material/printer knowledge, and constrained LLM reasoning for task-driven 3D printability assistance.Source: §8 · p. 15
• Experiments report improvements in material selection, physical printability, task suitability, and expert-rated report quality over rule-based, pure LLM, and LLM + KG baselines.Source: §8 · p. 15
• Post-print self-improvement evaluation suggests observed failures can refine subsequent recommendations, supporting interactive, task-aware printability assistance.Source: §8 · p. 15

Online Supplementary Material

• Representative interactive outputs and case demonstrations are available on the project page.Source: Online Supplementary Material · p. 15