PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhaoda Du, Qiaojie Zheng, Xiaoli Zhang

arXiv:2608.22128v1cs.AI

TL;DR

Functional 3D printing requires more than geometry-level printability, yet task suitability often depends on expert selection and post-print iteration. This paper grounds LLM reasoning in geometry evidence and structured material/process knowledge, improving material selection, physical printability, task suitability, and report quality.

  • Problem

    Geometry-level printability does not establish whether a part satisfies its intended use, while task suitability often depends on expert selection and repeated post-print evaluation.

  • Method

    The framework combines geometry evidence, an expert-validated Task-Material-Process knowledge graph, and constrained LLM reasoning to generate task-driven recommendations.

  • Results

    Experiments show improved material selection, physical printability, task suitability, and expert-rated report quality versus rule-based, pure-LLM, and LLM-plus-KG baselines.

  • Takeaways & Limitations

    Reliable task-driven printability assistance requires both geometry-grounded fabrication evidence and task-material-process knowledge.

  • Takeaways & Limitations

    Evaluation is limited to four focused STL models and four laboratory-available materials, requiring broader validation across geometries, printers, materials, and functional tests.

Abstract

from arXiv · show

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.

1. Introduction

The introduction frames printability and task suitability as distinct but coupled objectives that are often assessed sequentially, requiring expert intervention and repeated printing. It proposes a geometry- and knowledge-grounded LLM framework to jointly connect task intent with geometry, material, and process requirements before fabrication.

  • Motivation: Geometric printability determines whether a CAD or STL model can be fabricated, whereas task suitability determines whether the printed part meets its intended use.The two objectives operate at different levels and are coupled through material and process choices.
  • Problem: Existing automated methods detect geometry-related risks but do not infer intended use or jointly recommend materials and processes under task requirements.A model judged printable with PLA may still soften or deform when the task requires use under heat or other demanding conditions.
  • Problem: Task suitability commonly depends on expert material/process selection, post-print evaluation, and iteration, often requiring repeated printing before a suitable part is obtained.This workflow increases dependence on expert knowledge and delays identifying unsuitable choices until after fabrication.
  • Contribution: The proposed framework uses LLM reasoning to connect ambiguous task intent with geometry, material, and process requirements for joint pre-print assistance.It is designed to address the gap in jointly reasoning about printability and task suitability for a specific geometry and user task.
  • Framework: Given task intent and an STL model, the framework combines structured geometry evidence with a Task-Material-Process knowledge graph and constrained multi-objective LLM reasoning.The geometry evidence includes mesh-level computation, orientation scoring, heuristic risks, parameter baselines, and material candidates.
  • Evaluation: The evaluation covers multiple LLM backbones, novice-style task descriptions, same-geometry different-task cases, material selection, report quality, and self-improvement on selected problematic cases.Evaluation is organized around geometry-level printability, task suitability, material selection accuracy, and expert assessment of report quality.

2. Related Work

Prior printability-assessment methods provide geometry- and process-based manufacturability evidence but generally assume the setup and do not infer task intent. Data-driven and LLM-based approaches broaden prediction and interaction, yet existing systems remain limited in explanation, generality, or pre-print task-conditioned assistance from uploaded STL models.

  • Rule-based assessment: Traditional checkers compare extracted geometric features with printer constraints, including build volume, minimum feature size, wall thickness, overhang limits, support requirements, and mesh validity.These methods assess CAD or STL models under a predefined manufacturing setup.
  • Rule-based assessment: Setup-first methods identify geometry-related risks but usually do not infer user task intent or adapt material, parameters, and design recommendations to task suitability.Printer, material, orientation objective, and process constraints are assumed before assessment.
  • Data-driven assessment: Data-driven methods learn relationships between geometry, process conditions, and print outcomes using supervised learning, decision trees, neural networks, or STL-derived representations.Process-aware models can incorporate material behavior and parameters for geometric-quality prediction, inverse parameter selection, printability maps, process-window prediction, and defect monitoring.
  • Data-driven assessment: Data-driven and process-aware models remain limited for task-driven assistance because they often require substantial training data, provide limited decision explanations, and are tied to specific machines, materials, process windows, geometry families, or quality metrics.Their usefulness for prediction and parameter-selection guidance does not remove these generality and explanation constraints.
  • LLM-based assistance: LLMs support natural-language interpretation, reasoning over ambiguous goals, and human-readable explanations, but existing manufacturing systems mainly address querying, monitoring, or defect diagnosis rather than pre-print task-conditioned assistance from uploaded STL models.Prior work includes additive-manufacturing knowledge querying, defect prediction, process monitoring, and manufacturing decision support.

3. Methodology · 3.1. Geometry-Grounded Layer

The framework combines geometry/topology analysis, task cues, fabrication heuristics, and structured task–material–process knowledge to support downstream LLM reasoning. Its Geometry-Grounded Layer transforms STL geometry and lightweight task cues into evidence covering mesh validity, orientation, risks, design guidance, parameters, and material candidates.

  • 3. Methodology: The framework comprises three functional layers plus an optional post-print self-improvement module, integrating geometry evidence with structured task, material, printer, and process knowledge.The Geometry-Grounded Layer models task–geometry relations, while the knowledge graph links task requirements, material properties, and printer/process constraints.
  • 3.1. Geometry-Grounded Layer: The Geometry-Grounded Layer converts STL geometry and lightweight task cues into a structured evidence package for downstream LLM reasoning.It performs mesh-level computation, orientation scoring, and heuristic evidence generation and scoring.
  • 3.1. Geometry-Grounded Layer: The layer represents mesh, topology, orientation, heuristic, parameter, material-candidate, and orientation-selection outputs as components of its geometry-grounded evidence.The notation distinguishes mesh and topology evidence, orientation scores, heuristic risks, design hints, an initial parameter baseline, initial material candidates, and the selected orientation.
  • 3.1.1. Mesh-level Computation: Mesh-level computation parses the STL as a triangle mesh and calculates dimensions, surface area, estimated volume, and topology indicators relevant to slicing validity.Topology checks include watertightness, boundary edges, non-manifold edges, and degenerate triangles.
  • 3.1.2. Orientation Scoring: Orientation scoring evaluates six canonical build orientations using support-risk, base-contact, slenderness, and height-dominance indicators.The indicators are combined into a heuristic orientation score, and the lowest-scoring orientation is selected as the best candidate.
  • 3.1.2. Orientation Scoring: The orientation-score weights were empirically determined on primitive geometries to balance support burden, instability, bed-adhesion risk, and height dominance.The resulting score and selected orientation enter the evidence package for heuristic generation and LLM reasoning.
  • 3.1.3. Heuristic Evidence and Scores: Heuristic evidence generation converts geometry evidence and task cues into rule-triggered risks and design guidance, including warnings for support, adhesion, mesh validity, and stability.Task cues emphasize different concerns: appearance tasks prioritize support scarring, fit tasks prioritize dimensional accuracy and stability, and heat-related tasks prioritize material and shape retention.

3.2. Task-Material-Process Knowledge Graph · 3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer

The framework uses an expert-supervised Task-Material-Process knowledge graph and a reasoning layer that combines task descriptions with geometry- and graph-grounded evidence. This produces an auditable structured recommendation with constrained material and process guidance rather than unconstrained selection.

  • 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 feasibility.Its nodes include 8 application, 12 material-property, 7 printer/process constraint, and 25 candidate-material nodes; its edges include 72 application-material, 96 property-material, and 179 printer/material feasibility edges.
  • 3.2. Task-Material-Process Knowledge Graph: The KG was constructed through an expert-supervised, documentation-grounded, and LLM-assisted process under a fixed schema.The LLM organized information from official manufacturer material-printer documentation and assisted with grouping application and property categories and identifying missing task-property mappings.
  • 3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer: The reasoning layer receives task description T, geometry-grounded evidence E_G, and KG-based material evidence E_KG to generate the final structured recommendation report.The report covers printability, material choice, process parameters, design guidance, risks, and explanatory rationale.
  • 3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer: The report is auditable because its verdict, material recommendation, risks, process guidance, design guidance, and rationale trace to task intent, geometry evidence, or graph-retrieved knowledge.These outputs are represented as y, M, R_T, Θ, D_T, and E, respectively.
  • 3.3. Geometry- and Knowledge-Grounded LLM Reasoning Layer: Geometry-grounded evidence supplies task-aware geometry risks, orientation guidance, parameter baselines, and design hints, while KG evidence supplies ranked material candidates and supporting evidence traces.Jointly using T, E_G, and E_KG enables constrained task-driven reasoning.
  • 3.3.1. KG-Based Material Evidence Generation: The task-conditioned KG scoring procedure activates task-relevant application and property nodes, filters materials by printer/process compatibility, scores evidence and penalties, and ranks candidates.For each candidate, it retrieves evidence paths, computes application and property evidence scores, applies difficulty and feasibility penalties, stores supporting traces, and returns E_KG.
  • 3.3.1. KG-Based Material Evidence Generation: Material ranking uses task-activated application and property evidence together with material-difficulty and printer/process-feasibility penalties.Difficulty includes drying requirements, warping tendency, extrusion stability, and tuning sensitivity; printer/process infeasibility acts as a feasibility constraint under the fixed experimental setup.
  • 3.3.1. KG-Based Material Evidence Generation: Positive-scoring materials are passed to the LLM with supporting KG traces, preventing unconstrained material selection and integrating KG evidence with geometry-grounded evidence during report generation.This grounds final recommendations in task-conditioned material evidence and geometry evidence.

3.4. Post-Print Self-Improvement Module

The post-print self-improvement module optionally uses user feedback and stored print records to refine future recommendations through case-level retrieval and updates to geometry heuristics and the Task-Material-Process knowledge graph.

  • Feedback Collection: Users can provide post-print feedback on outcomes including success or failure, visible defects, dimensional fit, support removal, material behavior, or overall satisfaction.The feedback is collected after printing as an optional refinement mechanism.
  • Feedback Collection: The framework stores feedback with task intent, geometry evidence, generated report, selected material, and process settings.The stored process settings are denoted by Θ, while M denotes the material used.
  • Recommendation Refinement: Stored records refine recommendations through case-level memory and LLM-assisted updates to geometry-grounded heuristics and the Task-Material-Process knowledge graph.Records can be retrieved for similar tasks, materials, geometries, or failure modes, while geometry refinement can adjust risk-trigger thresholds and orientation-scoring preferences.

4. Experimental Design

The experiments use focused STL benchmarks paired with novice-style task descriptions to isolate printability challenges and test task-specific recommendations. Evaluations are conditioned on a fixed Bambu Lab X1 Carbon setup, four candidate materials, and controlled method/backbone comparisons.

  • Benchmark design: The benchmark includes overhang, dimensional accuracy, fine positive feature, and bridging models, each pairing geometric risks with task-specific descriptions.This design evaluates generated risks, explanations, and material, parameter, and design suggestions.
  • Task specification: Each focused STL benchmark is paired with two novice-style descriptions that state desired outcomes in non-technical language.Implicit task priorities and reference materials serve as ground-truth labels for evaluating intent inference and material recommendation accuracy.
  • Experimental conditions: All experiments use a fixed Bambu Lab X1 Carbon printer profile, so recommendations are conditioned on the available X1C setup rather than printer selection.The material candidate set is limited to PLA, TPU 95A, PETG, and PC.
  • Compared methods: Four method settings compare rule-based, pure LLM, LLM + KG, and the proposed geometry- and knowledge-grounded approach.The settings progressively add LLM reasoning, structured material knowledge, and geometry-grounded evidence.
  • LLM backbones: The methods are evaluated with Gemini 2.5 Flash-Lite, Flash, and Pro under the same backbone settings to isolate grounding effects from backbone choice.The comparison focuses on geometry and knowledge grounding rather than the choice of LLM backbone.

5. Evaluation Metrics Overview

The evaluation examines printability, task suitability, material-selection accuracy, expert-rated report quality, and self-improvement. It combines physical validation, repeated material predictions, blinded expert scoring, and post-print regeneration tests.

  • Printability evaluation: A trial is printability-successful only when the recommendation is executable and produces a physical print without a major geometry-level failure.Omitting essential information such as material, orientation, or process parameters counts as a printability failure.
  • Printability evaluation: 96 trials cover 8 task-conditioned scenarios, 4 methods, and 3 repetitions, with 75 physical prints and 21 non-executable recommendations counted as failures.Each method contributes 24 validation trials.
  • Task suitability evaluation: Task suitability is assessed only among successfully fabricated samples, which must pass the corresponding binary task-level test.Reported metrics are task-suitability among successful prints and the end-to-end task-suitable rate.
  • Material selection evaluation: Material-selection accuracy tests whether the primary recommendation matches the reference label across repeated task descriptions and method-backbone combinations.Each of eight task descriptions is evaluated five times, yielding 40 predictions per LLM-based method and 360 LLM-based predictions overall; the rule-based baseline adds 40 cases for 400 total.
  • Expert report evaluation: Expert evaluation scores six report components on usefulness for pre-print recommendations, using 10 randomly selected cases and three blinded experts evaluating anonymized reports from four methods.The rubric does not directly measure downstream functional performance.
  • Self-improvement evaluation: The proposed framework’s self-improvement module regenerates recommendations after adding post-print outcomes for 10 problematic cases and six failed physical prints.Regenerated outputs are evaluated using the main criteria, including material selection accuracy, printability, task suitability, and expert report quality.

6. Results and Analysis

The proposed framework delivers the strongest combined printability, task-suitability, material-selection, and report-quality results by integrating task-aware reasoning, geometry evidence, and structured knowledge. Post-print self-improvement further improves selected material recommendations and report quality.

  • Printability: The proposed framework achieves the best overall printability by combining task-aware material reasoning with geometry-grounded risk analysis and orientation guidance.This combination produces more complete and executable recommendations than the compared baselines.
  • Task suitability: 88.9% task-suitability among successfully printed samples (16/18) and 66.7% end-to-end task-suitable rate (16/24) are the highest reported results.The end-to-end rate compares with 33.3% for the rule-based baseline, 20.8% for Pure LLM, and 0.0% for LLM + KG.
  • Material selection: 90.0% material-selection accuracy for Gemini 2.5 Flash-Lite improves from 37.5% under pure LLM prompting.Adding the knowledge graph substantially improves accuracy, while the framework further improves upon Pure LLM + KG.
  • Expert evaluation: The framework achieves the highest mean expert score across all six structured report components, especially for material recommendation, risks, parameters, and explanations.These gains reflect stronger connections between task intent, geometry-dependent failure modes, and feasible slicer-level actions.
  • Post-print self-improvement: 7/10 problematic material recommendations are correct after self-improvement, compared with 0/10 before it, while expert scores improve across all six report components.A same-geometry different-task example also produces PETG for rigid fitting and TPU 95A for snap-fit use, illustrating task-conditioned recommendations.

7. Discussion

The discussion identifies geometry-grounded fabrication evidence and task-material-process knowledge as jointly necessary for task-driven printability assistance. It also highlights task-suitability trade-offs, evaluation limitations, and LLM inference as the main runtime bottleneck.

  • Core requirements: Task-driven printability assistance requires both geometry-grounded fabrication evidence and task-material-process knowledge.Geometry evidence makes recommendations physically executable.
  • Limitations and trade-offs: Materials satisfying functional requirements such as flexibility or heat resistance can introduce higher printing risks, creating a trade-off between task suitability and fabrication difficulty.The remaining failures reflect this trade-off.
  • Limitations and trade-offs: The evaluation is limited to four focused STL models and four laboratory-available materials, requiring broader validation across geometries, printers, materials, and downstream functional tests.The paper identifies broader validation as necessary.
  • Runtime: 45.871 s is the end-to-end runtime for a single Gemini 2.5 Flash-Lite request, with LLM inference as the main bottleneck.The geometry-grounded layer takes 0.003 s, KG-based material reasoning takes 5.350 s, and LLM inference takes 40.518 s; local geometry analysis adds negligible overhead.

8. Conclusion

The paper presents a task-driven 3D printability assistance framework combining geometry-grounded analysis, structured material/printer knowledge, and constrained LLM reasoning. Experiments indicate improved material selection, physical printability, task suitability, expert-rated report quality, and post-print recommendation refinement.

  • Framework: The framework combines geometry-grounded analysis, structured material/printer knowledge, and constrained LLM reasoning for task-driven 3D printability assistance.It is designed to connect printability assessment with task-oriented recommendations.
  • Evaluation: Experiments show improved material selection, physical printability, task suitability, and expert-rated report quality compared with rule-based, pure LLM, and LLM + KG baselines.The comparison spans both fabrication outcomes and the quality of generated reports.
  • Self-improvement: Post-print self-improvement evaluation suggests that observed failures can refine subsequent recommendations.This indicates that post-print feedback can support iterative recommendation improvement.

Online Supplementary Material

The project page provides representative interactive outputs and case demonstrations as supplementary material.

  • Representative interactive outputs and case demonstrations are available on the project page.
Loading 2608.22128v1…