Evolving Multimodal Models for Physical Dynamics: A Multi-objective Neuroevolution Approach

Accurate and efficient modeling of spatiotemporal physical dynamics is pivotal for a broad spectrum of applications, ranging from environmental monitoring to climate science. Nevertheless, the manual design of sophisticated deep learning architectures for this task remains a formidable challenge: the architectural and parametric design space is vast and highly entangled, frequently yielding suboptimal trade-offs among accuracy, efficiency, and generalization. To overcome this barrier, we propose a novel multi-objective neuroevolutionary framework for the automated design of physics-informed, multimodal models. The framework systematically navigates the joint architectural–parametric landscape to explicitly balance the inherent tension between predictive fidelity and computational cost. Concretely, it evolves candidate models that (i) exploit Large Language Models (LLMs) to interpret and embed textual physical constraints—e.g., governing equations and boundary conditions—into the learning process, and (ii) leverage manifold learning to uncover low-dimensional latent structures for principled computational reduction. Extensive experiments show that our framework automatically discovers a Pareto front of solutions that consistently surpass state-of-the-art baselines. The evolved models attain prediction errors as low as 0.0007 on the Diffusion–Reaction system and 0.2465 on the Navier–Stokes equations, while concurrently identifying configurations that cut computational overhead by up to 61% in FLOPs. These results demonstrate the framework’s capacity to deliver a suite of Pareto-optimal models tailored to diverse deployment requirements, charting a new direction for automated scientific discovery.