An AI system that repeatedly adjusted settings for selective laser sintering (SLS), the process used to make the polymer specimens, recorded stronger best-specimen results in later batches of PA11 Onyx and a PA12 blend, according to a new preprint. The study tested whether an “agentic” system — one that uses earlier build data, database memory and machine controls to choose new settings — could move polymer prints toward ASTM mechanical-property targets. The result is a feasibility signal on one platform, not a direct material-for-material test against manual tuning.
The work, identified as arXiv version 1 dated 26 August 2026, evaluated three materials on an Inova Mk1: PA12 GF, PA11 Onyx and a blend made from 25% PA12 GF and 75% PA12 White by volume.
The biggest changes came after repeated tuning
The clearest reported change was in PA11 Onyx. The best tensile modulus, a measure of stiffness, rose from 934 megapascals in Batch O to 1,571 MPa in Batch R. Batch R also recorded an ultimate tensile strength of 36.3 MPa, a flexural modulus of 1,093 MPa and a flexural strength of 42.5 MPa.
The early PA12 GF result set a low starting point. In the initial Batch A, the measured tensile modulus was around 365 MPa, compared with 2,800 MPa in the manufacturer’s technical data sheet. The study also includes a manually tuned PA12 GF control series spanning 15 batches. Because that series follows a different material and tuning track from the PA11 Onyx and PA12 Blend runs, it supplies context rather than a balanced head-to-head comparison.
For PA12 Blend, Batch U had the highest reported set of values: 1,678 MPa tensile modulus, 44.4 MPa ultimate tensile strength, 1,830 MPa flexural modulus and 60.6 MPa flexural strength. Its reported tensile strength was higher than the PA12 GF technical-data benchmark of 38 MPa, and its flexural strength was above the 56 MPa benchmark. Its tensile and flexural moduli, however, remained below the corresponding PA12 GF values of 2,800 MPa and 2,400 MPa.
One step in the blend sequence was especially notable: when the settings from Batch T were otherwise retained and the surface temperature was set to 172 °C for Batch U, ultimate tensile strength rose from 35.2 MPa to 44.4 MPa. The paper describes that difference as an increase of about 25%, although the study design does not establish that the temperature change alone caused it.
How the system worked
The system combined language-model reasoning with a PostgreSQL database for memory, Model Context Protocol tools and controls that could change machine parameters and apply runtime overrides. It used prior-build information, mechanical-test results and machine-level controls to select and revise settings over successive builds.
The tests followed ASTM D638 Type 1 and Type 4 procedures for tensile specimens and ASTM D790 for flexural testing. At least five samples per batch were planned. The builds kept the powder-chamber preheat at 145 °C, the layer height at 100 micrometres and the recoater speed at 100%.
Stronger numbers did not mean uniform parts
The mechanical results varied across the print bed. In PA11 Onyx Batches Q and R, stronger specimens tended to be near the recoater side, while weaker or insufficiently sintered specimens tended to be near the overflow side. That pattern means a best specimen may not represent every part made in the same build.
The PA12 blend showed some improvement in consistency: Batch U had a more even tensile-modulus distribution than earlier batches. But a property gradient remained across the print surface, so the process had not produced fully uniform mechanical behavior.
Batch O for PA11 Onyx yielded only three successfully obtained and tested samples because of significant warping, even though five samples per batch had been planned as a minimum. That means the Batch O best value rests on fewer tested samples than planned, and the PA11 sequence is reported through best-specimen results rather than a single uniform result for the whole batch.
A feasibility result, not a machine-wide verdict
Taken together, the results support a narrow claim: an agentic feedback loop was used for iterative SLS parameter optimization on the tested Inova Mk1 setup. That is different from showing that it is better than manual tuning, because the manual reference in the study was a PA12 GF series tuned over 15 batches rather than a balanced material-by-material comparison.
The numbers also should not be read as proof that all parts from a build had the same mechanical performance. PA11 showed bed-position differences, and the PA12 blend still had a property gradient across the print surface.
The study covered three materials on one Inova Mk1 platform, so its results are tied to that tested setup. For the PA12 blend, the external comparison was with manufacturer TDS values: the blend exceeded those benchmarks for strength but remained below them for modulus. Whether the same iterative process transfers to other machines or material lots remains open.
The supplied front matter gives a corresponding-author email but no funding or conflict-of-interest statement. The work remains an arXiv preprint, version 1, dated 26 August 2026.
Paper data and sources
Original title: AI Agentic Selective Laser Sintering Process Optimization
Authors: Peter Pak, Victor Alvarado, Amir Barati Farimani
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text