This work presents a novel framework leveraging Large Language Model (LLM)-enabled multi-agent systems to automate and accelerate the evaluation of alloys for Additive Manufacturing (AM). The traditional process of alloy selection and parameter optimization is complex, requiring deep expertise in materials science, thermodynamic simulations (like CALPHAD), and computational fluid dynamics (CFD). The proposed agentic system intelligently dispatches tool calls via protocols like Model Context Protocol (MCP) to perform sequential tasks: calculating thermophysical properties, simulating melt pool behavior, and generating process maps to identify defect-free parameter windows, specifically for lack-of-fusion defects.
The framework is built on a multi-agent LLM architecture where specialized agents reason through user prompts, plan task trajectories, and execute tool calls dynamically based on intermediate results.
The system employs a coordinator agent that decomposes a high-level query (e.g., "Evaluate SS316L for LPBF") into subtasks. Specialist agents then handle specific domains: a Thermodynamics Agent interfaces with CALPHAD software, a Process Simulation Agent calls solvers (Eagar-Tsai, Rosenthal, or OpenFOAM), and an Analysis Agent interprets results to generate process maps and recommendations. Communication and tool dispatch are standardized using MCP.
For a given alloy composition, the system automatically queries CALPHAD databases to compute equilibrium phases and temperature-dependent properties critical for AM simulation: thermal conductivity ($k$), specific heat capacity ($C_p$), density ($\rho$), and solidus/liquidus temperatures. This replaces manual database lookup and input preparation.
Using the material properties, the system executes analytical (Eagar-Tsai) or CFD (OpenFOAM) melt pool simulations across a range of beam power ($P$) and scan velocity ($v$) parameters. The resulting melt pool dimensions (width $w$, depth $d$) are used to calculate the lack-of-fusion (LoF) criterion. A process map is generated, delineating the "safe" parameter window from the defect-prone region.
The core of the defect prediction lies in melt pool modeling and overlap criteria. The Rosenthal solution for a moving point heat source provides a quick temperature field estimate: $$T - T_0 = \frac{P}{2 \pi k R} \exp\left(-\frac{v(R+x)}{2\alpha}\right)$$ where $T_0$ is ambient temperature, $R$ is radial distance from source, $v$ is scan speed, and $\alpha$ is thermal diffusivity. For LoF prediction, a critical condition is that the melt pool depth must exceed the layer thickness ($t$): $d \geq t$. For adjacent scan tracks, the overlap ratio $\eta = \frac{w_o}{w}$ (where $w_o$ is overlap width) must be sufficient, typically >~20%, to prevent voids.
The paper demonstrates the framework on two common AM alloys: Stainless Steel 316L and Inconel 718 (IN718). For each, the agent system was tasked with evaluating the standard composition and several proposed variants (e.g., IN718 with adjusted Nb content). The workflow involved: 1) CALPHAD calculation of liquidus temperature and $C_p$, 2) Eagar-Tsai simulation for a $P-v$ matrix (e.g., $P$: 50-300 W, $v$: 200-1500 mm/s), 3) Calculation of melt pool geometry, and 4) Generation of a 2D process map with LoF boundary.
The primary output is a Lack-of-Fusion Process Map. The chart is a 2D contour plot with Beam Power (W) on the Y-axis and Scan Velocity (mm/s) on the X-axis. A distinct boundary curve separates the chart into two regions. The lower-left region (low power, high speed) is shaded in red and labeled "Lack of Fusion Defect Region," where melt pool depth is insufficient. The upper-right region (higher power, moderate speed) is shaded in green and labeled "Stable Process Window." For IN718 variants, the map showed a measurable shift in the boundary curve, indicating that composition changes alter the optimal processing parameters. The agent system successfully quantified this shift and provided a comparative analysis.
~70%
Estimated reduction in manual setup & analysis time per alloy variant.
>500
Typical $P-v$ combinations simulated autonomously to map the defect boundary.
Example: Evaluating a Novel Al-Si-Mg Alloy Variant
User Prompt: "Assess the lack-of-fusion risk for AlSi10Mg with 1% increased Mg content for LPBF at a layer thickness of 30 µm."
This paper isn't about a new alloy or a better simulation solver; it's about orchestrating existing, siloed expert tools into a cohesive, autonomous workflow using LLMs as the "glue." The real innovation is the application of the agentic paradigm—inspired by frameworks like AutoGPT and Microsoft's TaskWeaver—to the notoriously iterative and multidisciplinary problem of AM alloy qualification. It directly attacks the bottleneck: the human expert's time spent translating between domain languages (materials, simulation, manufacturing).
The logic is compellingly sequential, mirroring an expert's thought process but automated: Composition -> Thermodynamics -> Properties -> Melt Pool Physics -> Defect Criteria -> Process Map. The use of lightweight analytical models (Rosenthal) for rapid screening before potentially invoking heavy CFD (OpenFOAM) shows intelligent resource allocation. This tiered approach is reminiscent of multi-fidelity modeling strategies used in aerospace design optimization.
Strengths: The system demonstrably accelerates the feedback loop for alloy evaluation. By leveraging LLMs' natural language interface, it lowers the barrier for materials scientists less familiar with simulation software. The dynamic task adjustment based on tool outputs is a key step toward robust autonomy.
Critical Flaws: The paper glosses over the "garbage in, garbage out" dependency on the underlying tools and databases. The accuracy of the final process map is wholly contingent on the CALPHAD database's fidelity for novel compositions and the limitations of the Eagar-Tsai model (which neglects fluid flow and keyhole dynamics). As noted in seminal CFD works like Khairallah et al., Physical Review Applied (2016), fluid flow can drastically alter melt pool geometry. An agent blindly trusting an analytical model could be confidently wrong. Furthermore, the evaluation is limited to a single defect (LoF), ignoring cracking, balling, and residual stress—a significant oversimplification of real-world AM challenges.
For industry adoption, the next step isn't just more agents; it's building validation feedback loops. The framework must integrate with experimental data (e.g., from in-situ monitoring like melt pool cameras or post-build CT scans) to calibrate and correct its simulations, moving towards a hybrid physical-AI model. Companies should pilot this on well-characterized alloys (like the SS316L shown) to benchmark its reliability before trusting it with novel materials. The ultimate vision should be a "Self-Correcting AM Advisor" that compares its predictions to real-world builds and continuously updates its internal models and recommendations.