researchsquare.comSource publication:
Loading content…
Design assessment for laser powder bed fusion no longer requires engineers to hand-carry results between preset rules and simulation tools. AgenticDfAM lets language-model agents interpret manufacturing guidance, select tools and propose redesigns, while geometric measurement and numerical comparison stay with executable tools. Across four case studies, three redesign strategies removed all reference violations on an analytical artifact, thermal feedback lowered a bracket's predicted layer-scale peak temperature by 160 °C, and a corrected support-construction procedure was reused in four build orientations.
Assessment and redesign are chained into one traceable agentic workflow in which agents revise both part geometry and the procedures used to assess and construct it. Prior design-for-additive-manufacturing tools applied predefined geometric checks, dimensional limits required experimental validation for specific alloys, machines and process settings, and engineers manually coordinated assessment, redesign and simulation. One rule catalog compiled from an IN718 guideline was applied to an analytical artifact, the CMU bracket, the NIST artifact and a clamp; the analytical artifact yielded 11 violations at a +Y build height of 110 mm but three at 70 mm in +Z or −Z, so compliance depended on build direction.
Numerical evidence stays with the tools, while agents handle applicability, explanation and strategy selection. Large language models alone cannot reliably provide precise geometric measurements, and errors in rule retrieval, dimensional interpretation and feature recognition have been documented. On the NIST additive manufacturing artifact, 70 of the 71 features specified in the drawing were recognized, plus 18 features present in the geometry but not separately specified, and all 226 quantities agreed with reference values at the reporting resolution; the central through bore was not recorded separately because its surface was paired with a coaxial outer surface as a tube.
Process feedback pushes redesign beyond geometric rules, and thermal analysis prompted an additional revision. The bracket had no reference-catalog violation, yet local stress and plastic-strain concentrations were predicted at the web-opening apexes and the web foot. The final T3 design kept all four openings and added 11.1% volume, lowering the predicted layer-scale peak from 537 to 377 °C while released displacement rose from 0.33 to 0.37 mm.
The support-construction procedure itself became the object of revision and was reused across four build directions. Initial support placed excess material above steep regions of curved faces even though the projected-footprint criterion was satisfied. After correction, support volume for the cylindrical target fell from 24,854 to 1,314 mm³ while the flat target stayed at 1,617 mm³ as a regression control; in +Z, support volume fell from 52,191 to 36,719 mm³ and the fraction with no part above it fell from 3.05% to 0.02%, and returns to support construction dropped from ten to five across four runs.
The demonstrated scope covers four geometries: an analytical artifact, the CMU bracket, the NIST artifact and a GrabCAD clamp. The rule catalog was compiled from one IN718 laser powder bed fusion guideline, with 15 entries across ten check types, 13 of them carrying executable numerical limits. Reuse of the support correction was shown on one clamp in four build directions, and the +Z design was rejected because its exported geometry measured a 1.20 mm flat unsupported span against a 1 mm limit. A reader interested in how guidance becomes executable checks with retained evidence can judge applicability from this scope.
Clamp displacement fields were treated as exploratory predictions because 0.2 mm support walls were not resolved by 1 mm mechanical voxels. Support-tooth dimensions were adopted from IN625 experiments without experimental validation on IN718, and support removability was not evaluated. In +Z, the in-loop solid measured a 1.00 mm unsupported span while the exported geometry measured 1.20 mm, and the operation responsible remained unresolved. Functional equivalence was not tested, so geometric compliance and feature retention do not establish it.