Researchers at Carnegie Mellon University (CMU) have developed a multi-agent, Large Language Model (LLM)-orchestrated workflow designed to automate early-stage alloy evaluation for laser powder bed fusion (LPBF). Reported in Additive Manufacturing Letters, the system integrates Thermo-Calc property calculations with an analytical melt pool model to generate lack of fusion (LoF) process maps for screening known and proposed alloy compositions.
Automating a multi-tool workflow for alloy screening
Selecting alloys for additive manufacturing typically requires coordination across thermodynamic databases, simulation tools, and defect modeling. In the CMU study, an LLM acts as an “agent” that dispatches tool calls to external software and interprets the results to evaluate alloy candidates and LPBF parameter windows.
The workflow combines three tool servers exposed through Model Context Protocol (MCP): a Thermo-Calc layer for CALPHAD-based property prediction, an additive-manufacturing package for process-map generation, and a workspace tool for state management. The LLM orchestrates these tools by converting natural-language prompts into structured inputs and summarizing outputs.
From thermophysical properties to lack of fusion maps
For each alloy, the system generates a composition file containing elemental mass fractions. Known alloys are retrieved from a lookup table, while hypothetical compositions are parsed from user prompts.
Thermo-Calc is used to compute density, thermal conductivity, specific heat capacity, and phase transition temperatures. A database-selection routine maps compositions to appropriate Thermo-Calc databases, with fallback options for multi-principal element alloys.
Absorptivity is estimated using a Drude-based approximation derived from electrical resistivity at 1070 nm. The authors note this serves as a baseline and may be less accurate for materials with strong power-dependent absorptivity effects.
The additive-manufacturing module then estimates melt pool dimensions using Rosenthal’s analytical heat-source model and applies a lack-of-fusion overlap criterion based on hatch spacing, layer height, and melt pool geometry. The resulting process map classifies beam power and scan velocity combinations as within or outside the LoF regime.
The current implementation assumes conduction-mode melt pool behavior and does not directly model keyholing or balling, which would typically require computational fluid dynamics (CFD) approaches.
Testing known, property-driven, and novel alloys
The system was evaluated across three scenarios: known alloys, property-driven searches, and novel or modified compositions.
For known alloys, including Stainless Steel 316L and Inconel 718, the workflow successfully generated LoF maps in most cases. In property-driven searches, the LLM proposed candidate alloys and compared their predicted LoF regimes.
Novel and modified compositions were also tested. While many prompts produced usable process maps, 2 of 10 novel alloy experiments failed due to thermodynamic constraints or database gaps, highlighting the practical limits of CALPHAD coverage.
Failure cases also emerged in known-alloy and property-based searches. A “Tool Steel” prompt stalled due to ambiguity between grade variants, and attempts to evaluate Mar-M 247 did not complete successfully. In another instance, the LLM incorrectly recommended Enhanced Maraging Steel despite it exhibiting the largest LoF region among candidates, illustrating how misinterpretation of tool-returned data structures can affect agent reasoning.
Validation against published LoF trends for IN718 and SS316L
The authors compared predicted LoF windows with published data for Inconel 718 and Stainless Steel 316L. For IN718, predicted LoF regions overlapped with literature observations, though some underprediction was attributed to the simplified Rosenthal-based model. For SS316L, predicted trends aligned with reported behavior, with LoF occurring at lower power and higher scan-velocity combinations.
Modeling limits and the broader shift toward integrated digital workflows
Keyholing and balling are discussed but not directly modeled, as capturing those regimes typically requires CFD methods that significantly increase computational cost. The authors plan to extend the framework to include additional defect regimes, beam-size effects, improved absorptivity modeling, and experimental validation of proposed compositions and parameter windows.
This roadmap reflects a broader constraint in metal additive manufacturing: high-accuracy thermal simulation is still computationally expensive, and alloy optimization workflows are often split across separate thermodynamic, modeling, and parameter-selection tools. Recent advances in LPBF modeling have focused on improving thermal simulation efficiency without relying on full-scale computational fluid dynamics, reflecting the persistent constraint that high-fidelity melt pool analysis is computationally expensive and slow to iterate. In parallel, industrial efforts toward full-stack metals optimization platforms have sought to integrate thermodynamic databases, process modeling, and performance targets into unified digital pipelines.
Within this context, the CMU agentic workflow extends that trajectory by introducing an orchestration layer that coordinates property calculation and process-map generation through natural-language prompts, while retaining physics-based solvers as the underlying decision engine rather than replacing them.
The 3D Printing Industry Awards are back. Make your nominations now.
Do you operate a 3D printing start-up? Reach readers, potential investors, and customers with the 3D Printing Industry Start-up of Year competition.
To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on LinkedIn.
While you’re here, why not subscribe to our Youtube channel? Featuring discussion, debriefs, video shorts, and webinar replays.
Featured image shows agentic workflow architecture. Image via Pak et al.
