image: This illustration shows how an AI-based “surrogate brain” is learned from neural data by combining forward modeling with inverse problem solving. Once trained, the surrogate brain can predict future large-scale brain activity, support dynamical systems analysis, enable virtual perturbation experiments, and guide model-based neurostimulation.
Credit: ©Science China Press
Artificial intelligence (AI) is transforming how scientists model and understand the brain. In a new review article, researchers from Southern University of Science and Technology, Huazhong University of Science and Technology, and Beijing Institute of Technology present a unified framework in which AI acts as a ``surrogate brain,'' capable of learning, predicting, and interpreting the brain's complex dynamics directly from data.
Traditional brain models rely heavily on fixed-form biophysical equations and population-averaged parameters. While mechanistically interpretable, such models struggle to capture the nonlinear, high-dimensional, and context-dependent activity of real neural systems. In contrast, modern AI---including recurrent neural networks, neural ordinary differential equations, graph-based architectures, and transformer models---can learn rich temporal and spatial patterns from large-scale neural recordings.
The review categorizes existing neural dynamical models into white-box (mechanistic), black-box (data-driven), and gray-box (hybrid) approaches. By integrating these perspectives, the authors propose a surrogate brain framework that covers three interconnected components: (i) constructing neural dynamical models, (ii) solving the inverse problem to infer parameters and latent states, and (iii) evaluating predictive performance and functional fidelity.
The authors highlight that surrogate brains enable multiple applications: predicting future whole-brain activity, performing dynamical systems analysis, conducting virtual perturbation experiments, and guiding neurostimulation strategies. These capabilities make surrogate brains valuable tools for both basic neuroscience and translational neuroengineering.
A major challenge discussed in the review is the ill-posed nature of inverse problems in brain modeling, including issues of existence, uniqueness, and stability of solutions. The authors summarize mathematical and neuroscience-informed regularization techniques---such as sparsity constraints, anatomical priors, physical-law constraints, and stability-enhancing optimization---to ensure solutions remain reliable even in the presence of noise and limited data.
The researchers envision that AI-based surrogate brains will play an increasingly central role in future neuroscience, serving as personalized computational counterparts of the human brain. Such models could one day support real-time prediction, individualized diagnostics, adaptive neurostimulation, and virtual experimentation.
"This framework provides a systematic bridge between theoretical neuroscience and clinical applications,'' the authors conclude. "AI-based surrogate brains offer a powerful path toward individualized, interpretable, and predictive models of brain function.''