Machine learning-based neural operators offer transformative potential for accelerating high-entropy alloy simulations while capturing complex many-body interactions beyond classical potentials. We are working on developing physics-informed neural networks (PINNs) and deep potential molecular dynamics (DeePMD) models trained on our FeNiCrCoAl compression data to predict stress-strain responses, defect evolution, and phase transformations at computational costs orders of magnitude lower than traditional MD. Graph neural networks could enable real-time prediction of local atomic environments during extreme deformation, and transfer learning approaches may generalize across compositional spaces, enabling rapid screening of novel HEA formulations for specific mechanical properties without extensive ab-initio calculations.

This study employs molecular dynamics simulations using LAMMPS to investigate laser-based manufacturing processes in aluminum alloys. We simulate two distinct processes: laser welding of butt-jointed aluminum plates and laser melting of spherical aluminum powder particles on a substrate. The simulations utilize the EAM potential to capture realistic melting behavior, incorporating localized laser energy deposition, melt pool dynamics, and thermal gradients. We analyze the evolution of temperature fields, stress distributions, and microstructural changes through coordination number and centro-symmetry parameters. The research provides atomistic insights into phase transformations, solidification patterns, and defect formation during laser-material interactions, contributing to the optimization of additive manufacturing and welding processes.

SolidAdditive AI is a specialized artificial intelligence assistant designed exclusively for solid-state additive manufacturing processes. Unlike conventional AI tools that attempt to cover all manufacturing domains, this application maintains laser-focused expertise on four critical solid-state processes: Cold Spray Additive Manufacturing (CSAM), Ultrasonic Additive Manufacturing (UAM), Friction Stir Additive Manufacturing (FSAM), and Additive Friction Stir Deposition (AFSD).

Built on Google's Gemini 2.5 Flash model, the application features a six-layer enforcement system ensuring 99.9% accuracy in maintaining its solid-state scope, automatically rejecting queries about fusion-based or polymer additive manufacturing. The system offers five specialized analysis modes, i.e., General, Microstructure Analysis, Process Design, Troubleshooting, and Comparison, each optimized with expert-level prompts tailored to specific use cases.

Application 

Additive Friction Stir Deposition (AFSD) builds metal parts layer by layer without melting the material. Instead of liquefying metal like traditional 3D printing or welding, AFSD uses rotation, pressure, and friction to smear solid material onto a surface. The material gets hot but stays solid, which avoids many problems associated with melting and re-solidifying metals. To understand what actually happens to the atoms during this process, we can run molecular dynamics simulations using LAMMPS software. We developed the first atomistic simulation-based approach to model the Additive Friction Stir Deposition (AFSD) process. 

The simulation starts with two main parts: a flat aluminum substrate at the bottom and a cylindrical aluminum rod above it. The substrate represents either the base plate or a previously deposited layer. The bottom of the substrate is frozen in place, so the whole system doesn’t just drift around during the simulation. Everything else can move freely and respond to the forces we apply. This setup mirrors what you would see in a real AFSD machine, i.e., a spinning rod of material being pressed down onto a surface.

Modern lattice structures contain hundreds of thousands of geometric points. The STL file used in this study contains 368,768 sampled points representing a periodic strut-based architecture. Working with this level of detail becomes computationally expensive for tasks like shape optimization, structural analysis, or design space exploration. Traditional mesh simplification methods often lose critical topological information or create artifacts that compromise the structural integrity insights we need.

This research applies Self-Organizing Maps, an unsupervised neural network, to learn a compressed representation of the lattice geometry. The SOM creates a 20×20 grid of nodes that organize themselves to match the shape and topology of the input structure. Think of it as teaching a flexible mesh to wrap itself around the complex geometry, learning where the important features are and how they connect. 

We have developed a comprehensive LAMMPS-based molecular dynamics simulation script to model the Friction Stir Welding (FSW) process at the atomic scale. The simulation captures the complete welding cycle, including tool plunge, traverse along the joint line, and retraction, while tracking critical parameters such as temperature evolution, material flow patterns, atomic coordination, and microstructural changes in the weld zone. Our script models the interaction between a rotating tungsten tool and aluminum workpieces, generating rich datasets that include atomic positions, velocities, temperature distributions, and defect formation throughout the welding process.

Building on this simulation framework, we plan to integrate machine learning and deep learning algorithms to unlock deeper insights and predictive capabilities. The extensive trajectory data generated from our simulations will serve as training datasets for neural networks to predict optimal welding parameters, identify defect formation patterns, and correlate process conditions with final weld quality. We envision using convolutional neural networks (CNNs) to analyze microstructural evolution from simulation snapshots, recurrent neural networks (RNNs) to predict time-series behavior of temperature and stress fields, and reinforcement learning algorithms to optimize tool path and rotation speed for different material combinations.

We recently carried out a simulation study on CoCrFeMnNi high entropy alloy casting using OpenFOAM. This equiatomic five-element alloy has gained significant attention in materials engineering due to its exceptional mechanical properties, and understanding its solidification behavior is crucial for optimizing manufacturing processes.

We modeled a rectangular casting mold with realistic thermal boundary conditions to replicate actual casting scenarios. The bottom surface was set to a cold temperature to simulate contact with a chill plate, while the side walls were moderately cooled to represent typical mold conditions. The top boundary remained insulated like a standard riser setup. Starting with superheated molten CoCrFeMnNi, we tracked both heat transfer and fluid flow throughout the entire solidification process using the buoyantPimpleFoam solver coupled with the solidificationMeltingSource model.