Source-linked AI summary
AI Agentic Selective Laser Sintering Process Optimization
Peter Pak, Victor Alvarado, Amir Barati Farimani
TL;DR
This work addresses the tedious, iterative challenge of selecting optimal SLS process parameters. It uses an agentic system to tune parameters for tensile and flexural properties across materials, achieving properties that exceed reference samples on PA12 Blend over a small number of iterations, while remaining constrained by Inova Mk1 limitations.
Problem
Optimal process parameters are tedious to determine because each parameter set requires multiple fabrication, testing, and analysis cycles, while hardware variation limits simple use of recommended settings.
Method
An agentic system uses calibration and tuning across PA12 GF and PA11 Onyx to optimize SLS process parameters for tensile and flexural properties.
Results
Mechanical properties exceeded those of reference samples on PA12 Blend, with process parameters optimized over a small number of iterations.
Takeaways & Limitations
The results support agentic tuning of SLS process parameters using knowledge from previous builds and limited user guidance.
Takeaways & Limitations
The Inova Mk1 has software and hardware limitations, and its comparatively coarse sensors may constrain data quality.
Abstract
from arXiv · showhide
Agentic systems enable the intelligent automation of complex workflows, specific to additive manufacturing this is applicable for complex tasks such as process parameter optimization for mechanical properties. This work investigates the AI enabled agentic process optimization within Selective Laser Sintering (SLS) to iteratively improve the tensile and flexural properties of 3 different materials on the Inova Mk1. These materials include PA12 GF, PA11 Onyx, and PA12 Blend (volume mixture of 25% PA12 GF and 75% PA12 White) and with using knowledge from previous builds and minimal guidance from the user, the agentic system was able to optimize process parameters over a small number of iterations to achieve comparable TDS specified mechanical properties. This work showcases the ability for an agentic system to continually learn from updated data, enabling the intelligent automation of complex tasks such as process parameter optimization for selective laser sintering.
1 Introduction
Selective laser sintering requires precise, machine-specific process parameters, making optimization laborious. This work introduces a human-in-the-loop agentic system that uses prior build data and tool-enabled reasoning to optimize parameters for mechanical properties across multiple materials.
- Process optimization challenge: Recommended settings cannot be transferred simply between environments and machines because process nuances differ.
- Agentic approach: Agentic systems use tools and data resources to let LLMs operate in dynamic environments and orchestrate complex workflows.Tools may include software platforms, APIs, and system-level commands.
- Process optimization challenge: SLS process parameters such as chamber temperature, laser power, and scanning velocity affect part quality.
- Process optimization challenge: Optimal parameter search is tedious because hardware variability, including optics, complicates calibration.A single parameter set requires multiple fabrication, testing, and analysis cycles.
- Study contribution: The study introduces an agentic SLS parameter-optimization system evaluated against ASTM standards across multiple materials.
- Study contribution: The human-in-the-loop system combines previous ASTM build data, firmware-level tool calls, and LLM reasoning to determine parameters for desired mechanical properties.Experiments use the open-source SLS4All Inova Mk1, whose collected optical, temperature, and positional data support continual learning.
2 Related Work
Prior work applies agentic systems to additive-manufacturing monitoring, process mapping, defect correction, and materials-science workflows. These studies demonstrate tool use, multi-agent orchestration, and data-driven reasoning across manufacturing and materials tasks.
- Additive manufacturing agents: LLM-3D Print uses multi-agent orchestration and image-based monitoring to detect layer defects and change parameters in subsequent layers.The system uses GPT-4o-orchestrated agents for information gathering, planning, and execution.
- Additive manufacturing agents: LLM-3D Print achieved compression-test maximum peak loads 5x greater than samples constructed without the system.
- Process mapping: Agentic alloy evaluation uses tool calling with Thermo-Calc and a Rosenthal-based thermal model to generate lack-of-fusion process maps.
- Process mapping: For known alloys, the alloy-evaluation system aligned well with literature process regimes across layer heights.
- Process mapping: For unknown compositions, predicted lack-of-fusion regimes followed the general defect trends expected for several alloy compositions.
- Materials-science agents: MatSciAgent combines four agents for materials retrieval, continuum simulation, crystal-structure generation, and molecular-dynamics simulation.GPT-3.5-Turbo serves as the core orchestrator in the reported case studies.
- Materials-science agents: AtomAgents uses physics-based simulation tools across materials-property calculation, dislocation analysis, multiscale mechanics, and hypothesis validation.
- Materials-science agents: A GNN-enabled agentic system searches NbMoTa alloy candidates using Peierls barrier and screw dislocation energy as optimization properties.
3 Methodology
The methodology uses an open-source Inova Mk1 SLS machine, sensor suite, and software stack to perform polymer powder processing and support agentic communication. It also documents hardware, software, and sensing limitations that constrain monitoring and material scope.
- 3.1.1 SLS4All Inova Mk1: The selective laser sintering process uses the Inova Mk1, an open-source, low-cost machine developed by SLS4All founders Tomas Starek and Pavel Dyntera.The machine was purchased as a kit and assembled over several months.
- 3.1.1 SLS4All Inova Mk1: The Inova Mk1 supports polymer powders including PA12, PA12 GF, and other experimental compositions.The manufacturer suggests primarily polymer-based powders with melting temperatures of 200 °C.
- 3.1.1 SLS4All Inova Mk1: The machine uses a 450 nm blue diode laser delivering 10 watts, a 150 mm x 150 mm x 185 mm build volume, and separate surface and chamber heating elements.Surface heating uses four halogen lamps, while chamber temperature control uses five heating elements.
- 3.1.1 SLS4All Inova Mk1: Average scan speed is around 1,650 mm/s at 5 watts with a 250 µm spot size and 2,800 mm/s at 10 watts with a 350 µm spot size.These are the two reported power, scan-speed, and spot-size combinations.
- 3.1.2 Platform Limitations: Hardware control runs through open-source software including Klipper and SLS4All.Compact, while custom firmware plugins provide the primary communication method for the agentic system.The software stack controls galvometers, the laser, stepper motors, sensors, and heating elements.
- 3.1.2 Platform Limitations: The platform has coarse sensing relative to other in-situ monitoring systems, potentially constraining the real-time information available while testing parameters and executing builds.The stated comparison concerns the resolution and frame rate of the installed thermal and optical imaging equipment.
- 3.1.2 Platform Limitations: Proprietary compiled slicing and tool-path modules, limited API documentation, and a dashboard streamed from the host constrain software transparency and integration.The maintainers nevertheless offer custom firmware plugins.
3.2 Agentic System
The agentic system automates process-parameter selection and adjustment by combining an LLM with tools, database memory, and machine-level controls. It uses prior build data and patch testing to investigate parameter–property relationships and guide subsequent experiments.
- Agentic System: The system enables automated selection and adjustment of process parameters for ASTM builds and runtime overrides.Overrides can address surface temperature, recoating, and applied energy.
- Agentic System: PostgreSQL stores prior outputs and tool-call results so the LLM can recall them as reasoning context for future prompts.
- MCP Enabled Tool Calling: MCP tools connect the LLM to material properties, previous builds, print profiles, ASTM data, and machine firmware functions.
- Agent Harness: The modular architecture allows tools and functionality to be used interchangeably across agent harnesses and supports replacing the current LLM.
- MCP Enabled Tool Calling: A 5 x 5 patch grid rapidly tests parameter-specific configurations, while sensors extract patch-quality information to select the next configuration.
- MCP Enabled Tool Calling: Recorded data tools let the LLM investigate previous parameter settings and apply adjustments intended to improve ASTM-measured mechanical performance.
3.3 Manufacturing Process
Manufacturing used prescribed job templates and largely fixed build conditions on the Inova Mk1, while batch-specific process parameters varied between experiments. Cooling, powder handling, and realized surface temperature introduced practical process constraints.
- Manufacturing Process: Each build used a 145 °C powder-chamber preheat, 100 µm layer height, 100% recoater speed, and a prescribed cooling procedure.
- Manufacturing Process: The Inova Mk1 firmware maintained print parameters, including halogen-lamp control intended to keep the print surface temperature even.
- Manufacturing Process: Realized surface temperature may differ from the target because halogen lamps are pulsed during operation.
- Manufacturing Process: Builds cooled overnight, often for 8+ hours, before powder-cake removal and postprocessing.
- Manufacturing Process: Powder was sifted and reclaimed during postprocessing, with materials kept separated to the author’s best ability.
3.4 Sample Evaluation
Mechanical-property evaluation followed ASTM specimen standards and compared measured tensile and flexural properties with manufacturer-published values. Samples shared fixed orientation and baseline conditions, with at least five specimens printed per batch.
- Sample Evaluation: Specimens were designed and tested according to ASTM D638 for tensile properties and ASTM D790 for bending stress analysis.
- Sample Evaluation: At least 5 samples were printed for each batch to obtain a range of benchmarking values.
- Sample Evaluation: Testing used an Instron 4469 Universal Testing System with a 50 kN load cell and the prescribed standard methodologies.
4 Results
Results show progression from manually tuned PA12 GF baselines to agentically optimized PA11 Onyx and PA12 Blend builds. Across the experiments, subsequent parameter changes improved measured mechanical properties, with PA12 Blend reaching several values comparable to or above reference and TDS values.
- Material 1 (PA12 GF): PA12 GF reference samples measured a tensile modulus of around 2599 MPa and a flexural modulus of 1950 MPa.The reference samples had spreads of 98 MPa and 271 MPa, respectively.
- Material 1 (PA12 GF): Batch H reached a tensile modulus of 2815 ±4 MPa and a flexural modulus of 2272 ±396 MPa, but these properties were not reproducible.
- Material 2 (Nylon 11 Onyx): The PA11 Onyx results indicate that the agentic system learned from prior batches and improved mechanical properties with subsequent prints.
- Material 3 (Nylon 12 Blend): PA12 Blend achieved an ultimate tensile strength of 44.4 MPa, approximately 25% above Batch T’s 35.2 MPa, alongside a flexural modulus of 1830 MPa and flexural strength of 60.6 MPa.
- Material 3 (Nylon 12 Blend): PA12 Blend’s flexural strength of 60.6 MPa exceeded both PA12 White and PA12 GF reference and TDS values.
5 Discussion
Reference samples and optimized batches showed that mechanical properties varied across the print bed and sometimes differed from manufacturer specifications. Recoater-side placement generally corresponded to higher tensile modulus, while Batch U produced the most even distribution among the Nylon 12 Blend batches.
- Reference-property comparison: PA12 White and PA12 GF technical data sheets reported greater mechanical-property values than reference samples obtained through testing.
- Reference-property comparison: PA12 GF reference samples differed from manufacturer specifications by up to 30%, with print orientation identified as one potential explanation.The deviations were minor for PA12 White but significantly greater for PA12 GF.
- Print-bed variation: PA11 Onyx batches showed higher tensile modulus values near the recoater, consistent with a possible chamber-temperature gradient.The passages associate higher temperatures with the recoater side of the print chamber.
- Print-bed variation: Mechanical properties decreased from the recoater side toward the overflow side in both ASTM D638 sample stacks.In Batch Q, overflow-side samples showed insufficient sintering, including one sample that broke before testing.
- Nylon 12 Blend batches: Batch U produced the best Nylon 12 Blend results and a more even tensile-modulus distribution, although a property gradient remained visible.Batches S and T also showed mechanical-property trends associated with placement along the print surface.
6 Conclusion
The agentic system tuned SLS process parameters using calibration on PA12 GF, further tuning with PA11 Onyx, and knowledge from previous builds. Across three materials, it achieved comparable TDS-specified properties over a small number of iterations while continually learning from updated data.
- The agentic system tuned SLS parameters for tensile and flexural properties across PA12 GF, PA11 Onyx, and PA12 Blend.
- The system used knowledge from previous builds and minimal user guidance to optimize process parameters over a small number of iterations.
- The optimized parameters achieved mechanical properties comparable to TDS specifications for various materials.
- The work demonstrates continual learning from updated data for intelligent automation of SLS process-parameter optimization.
Appendix A Manufactured Samples
The initial Nylon 12 Blend tensile samples appeared relatively lighter in color than samples made from PA12 GF or PA11 Onyx.
- Initial Nylon 12 Blend tensile samples had a relatively lighter color than PA12 GF and PA11 Onyx samples.The blend consisted of 25% PA12 GF and 75% PA12 White by volume.