Researchers at MegaOne AI, led by Senior Staff Scientist Dr. Anya Sharma, have introduced AutoMS, a novel multi-agent evolutionary search framework designed for cross-physics inverse microstructure design. The framework, detailed in a paper published on arXiv on March 27, 2026, addresses the complex challenge of designing microstructures that simultaneously meet multiple, often conflicting, physical objectives across different domains, such as mechanical and thermal properties. The work aims to accelerate the discovery of advanced materials with tailored functionalities by navigating the vast and discontinuous search spaces inherent in inverse design problems. The full paper is available here.
AutoMS employs a multi-agent system where individual agents explore different regions of the design space, iteratively refining microstructure geometries based on their performance across various physics simulations. This distributed approach allows for a more efficient exploration compared to traditional single-objective optimization methods, which often struggle with the high dimensionality and non-convexity of cross-physics design problems. The framework integrates a surrogate modeling component to reduce the computational cost associated with repeated high-fidelity physics simulations.
The core of AutoMS lies in its evolutionary search algorithm, which leverages principles of natural selection to evolve a population of microstructures. Each microstructure is evaluated against a set of predefined cross-physics objectives, and those exhibiting superior performance are selected for reproduction and mutation, generating new design candidates. This iterative process allows the system to progressively converge towards optimal or near-optimal solutions.
In a demonstration, AutoMS was applied to a problem involving the simultaneous optimization of a material’s stiffness and thermal conductivity. The framework successfully identified microstructures that achieved a 15% improvement in the combined objective function compared to designs generated by conventional single-objective topology optimization methods. The search process involved evaluating approximately 5,000 unique microstructure designs over 200 generations, demonstrating its capability to explore a substantial design space.
Furthermore, the researchers reported that AutoMS exhibited a 2.5x speedup in convergence time when compared to a baseline multi-objective genetic algorithm that did not incorporate the multi-agent and surrogate modeling components. This efficiency gain is attributed to the framework’s ability to intelligently prune less promising design regions and focus computational resources on areas with higher potential for improvement. The framework’s modular design allows for the integration of various physics solvers and objective functions, enhancing its adaptability to different material design challenges.
The current iteration of AutoMS primarily focuses on 2D microstructure design problems. Future work will involve extending the framework to handle complex 3D geometries and incorporating uncertainty quantification to account for manufacturing tolerances and material variability.
