Session: AM9.1 - Materials for (strategic sectors) Aerospace and Defense applications

Paper Number: 187552

187552 - Inverse Design of Mechanical Metamaterials Using Neural Surrogates and Meta-Heuristic Optimization 

Mechanical metamaterials provide an engineering route to achieve stress-strain responses that cannot be realized using conventional materials, including controlled stress plateaus, delayed densification, and high specific energy absorption (SEA). These behaviours are particularly valuable for defence and aerospace structures subjected to impact and blast loading, where survivability depends on the ability to dissipate energy while maintaining bounded transmitted stress. However, metamaterial development is still dominated by forward design, where unit-cell geometries are repeatedly modified and re-simulated until the response approaches the target. This workflow becomes impractical when performance requirements are specified directly in terms of stress-strain constraints, such as a mandated plateau stress range and a minimum energy absorption. The difficulty arises because the mapping from geometric design parameters to nonlinear mechanical response is highly coupled and non-intuitive, leading to costly redesign loops and limited exploration of the design space. This paper presents an artificial intelligence-enabled inverse design framework that directly predicts mechanical metamaterial geometry parameters from desired stress-strain targets for energy-absorption applications.

A high-fidelity dataset from finite element analysis (FEA) simulations of parametric metamaterial unit cells spanning multiple architectures and a wide range of geometric descriptors is utilized for training the neural surrogates. Each sample consists of a design vector and the corresponding stress-strain curve sampled on a fixed strain grid, enabling learning of the complete nonlinear response rather than isolated scalar properties. A deep neural surrogate model is trained to approximate the forward mapping from geometry to stress-strain behaviour. The surrogate is designed to capture the evolution of the curve across the elastic regime, the nonlinear transition, the plateau region, and densification. The trained model demonstrates strong predictive performance, achieving R² values exceeding 0.97 on the training set and 0.96 on the test set, confirming its ability to generalize across unseen geometric combinations and accurately reconstruct nonlinear response trends.

Inverse design is formulated as a constrained curve-matching problem in which the design parameters are treated as decision variables and the objective is to minimize the discrepancy between a user-defined target stress-strain curve and the surrogate-predicted curve. Meta-heuristic optimization is adopted due to the non-convexity of the objective landscape, the presence of multiple feasible solutions, and the need to enforce manufacturability and design bounds. Several meta-heuristic algorithms are implemented and benchmarked under identical target curves, including energy-absorption-driven targets characterized by (i) bounded peak stress, (ii) prescribed plateau stress level, and (iii) delayed densification strain. Algorithm performance is compared in terms of curve error, SEA error over specified strain windows, convergence robustness, and computational time. The predicted parameter sets are validated by re-evaluating the corresponding geometries using the forward surrogate and by quantifying error in response-defining features such as initial stiffness, plateau stress, and densification onset.

Results show that the proposed AI-based inverse design pipeline can consistently recover feasible metamaterial designs whose predicted responses closely match target stress-strain behaviour, including plateau-dominated curves required for efficient energy absorption. Compared to brute-force FEA search, the method reduces the number of expensive simulations by orders of magnitude while maintaining high response fidelity. Beyond design automation, the framework supports engineering interpretation through feature attribution analysis, which identifies the geometric parameters most responsible for governing plateau formation, stress localization trends, and densification behaviour.

The presented methodology provides a scalable inverse design route for mechanical metamaterials intended for defence and aerospace protective structures, including blast/impact mitigation panels, lightweight crashworthy cores, and sacrificial energy-absorbing layers. The framework is extensible to additional constraints such as multi-loading conditions, strain-rate effects, or manufacturing-aware objectives, enabling rapid synthesis of mission-specific metamaterial responses.

Presenting Author: Dheeraj S Nair Indian Institute of Information Technology, Design and Manufacturing, IIITDM Kancheepuram

Presenting Author Biography: Dheeraj S Nair is currently PhD student in the School of Interdisciplinary Design and Innovation at the Indian Institute of Information Technology, Design and Manufacturing, IIITDM Kancheepuram, Chennai. His research work focuses on auxetic metamaterial structures, AI in design, inverse design and deep learning approaches for architectured metamaterials.