Artificial intelligence as a surrogate brain: Bridging neural dynamical models and data

TL;DR

The paper foregrounds a unified AI-based surrogate-brain framework to bridge theoretical neuroscience and translational neuroengineering by predicting whole-brain dynamics from historical data through forward modeling, inverse problem solving, and rigorous evaluation. It systematically categorizes forward models into white-box, black-box, and gray-box families, and analyzes inverse problems via Bayesian and deterministic lenses, including well-posedness and regularization strategies. A comprehensive evaluation scheme combines mathematical metrics (e.g., $\text{MSE}$, $\text{EV}$, $\text{KL}$, Wasserstein distances) with neuroscientific measures (spectral content, functional connectivity, decoding) and introduces a multi-metric surrogate-brain benchmark. The framework supports applications in system analysis, in-silico simulation (e.g., VEP, TVB), and model-guided neuromodulation, while outlining challenges in multi-scale data integration, model-structure design, generalization/personalization, and ethics. Overall, the work provides a path toward accurate, interpretable, and personalized surrogate brains capable of aiding neuroscience research and clinical neuroengineering.

Abstract

Recent breakthroughs in artificial intelligence (AI) are reshaping the way we construct computational counterparts of the brain, giving rise to a new class of ``surrogate brains''. In contrast to conventional hypothesis-driven biophysical models, the AI-based surrogate brain encompasses a broad spectrum of data-driven approaches to solve the inverse problem, with the primary objective of accurately predicting future whole-brain dynamics with historical data. Here, we introduce a unified framework of constructing an AI-based surrogate brain that integrates forward modeling, inverse problem solving, and model evaluation. Leveraging the expressive power of AI models and large-scale brain data, surrogate brains open a new window for decoding neural systems and forecasting complex dynamics with high dimensionality, nonlinearity, and adaptability. We highlight that the learned surrogate brain serves as a simulation platform for dynamical systems analysis, virtual perturbation, and model-guided neurostimulation. We envision that the AI-based surrogate brain will provide a functional bridge between theoretical neuroscience and translational neuroengineering.

Figures (5)

• Figure 1: Conceptual framework for AI-based surrogate brains in neural dynamical systems. A surrogate brain is constructed through two interconnected processes: forward modeling and inverse problem solving. Forward modeling describes how latent brain states $\mathbf{x}_t$ evolve according to a dynamical operator $F_\theta(\cdot)$ and produce observable signals via an observation mapping $h_\phi(\cdot)$. Here, $\mathbf{u}_t$ represents external inputs, $\boldsymbol{\xi}_t$ captures intrinsic dynamical noise, and $\boldsymbol{\eta}_t$ models measurement uncertainty. Inverse problem solving entails learning $F_\theta(\cdot)$ and $h_\phi(\cdot)$ from data, which involves selecting training strategies (e.g., next-token prediction), defining objectives, and optimizing parameters. The framework accommodates white-box, gray-box, and black-box formulations of $F_\theta(\cdot)$ and $h_\phi(\cdot)$, enabling the surrogate brain to flexibly integrate mechanistic priors with data-driven adaptation. Together, these steps yield a personalized, predictive model of brain dynamics that supports mechanistic insight, virtual experimentation, and model-guided neurostimulation.
• Figure 2: Three types of neural dynamical models: white-box, black-box, and gray-box models. A. White-box models at multiple scales. Left: biophysical single-neuron models; center: a neural-mass model obtained by mean-field reduction of many single neurons; right: a whole-brain network in which each node is a neural-mass model. Neural mass model in panel A is an original schematic adapted from the conceptual framework in Ref. breakspear2017dynamic. B. Black-box models. Top: modelling directly in the observation space, where a neural network predict the future neural dynamics, $\mathbf{\dot{x}}=\hat{f}(\mathbf{x},\theta)$; bottom: modelling in a latent space, where the data are first encoded/decoded and dynamics are learned as $\dot{\mathbf{z}}=\hat{f}(\mathbf{z},\theta)$. C. Gray-box models incorporating priors. Top: neural dynamics models constrained by neuroscientific priors (including dend-PLRNN, reproduced from Ref. brenner2022tractable under a CC BY 4.0 license), E-I RNN song2016training, Low-rank RNN pellegrino2023low and Spiking Neural Network (SNN)); bottom: dynamical models constrained by physical-law priors(including Physics-Informed Neural Network (PINN) and Hamiltonian Neural Network (HNN)).
• Figure 3: Solving inverse problem. A. Modelling.B. Learning with regularization. Priors on object attributes are added as regularization to the loss function, and the optimizer solves the learning problem. C. Example of non-uniqueness. Different initializations in the optimization algorithm yield distinct solutions with comparable model performance. D. Example of instability. Stability analysis via singular value spectra (top) and perturbation testing (bottom).
• Figure 4: Model evaluation from mathematical and neuroscientific perspectives.A–C| Mathematical perspective. Synthetic data were generated with the stochastic Jansen–Rit model Ableidinger2017ASV. The upper and lower rows correspond to two surrogate brains derived from the same underlying system, evaluated across three representational spaces: A. state-space trajectories(mean root-mean-square error, $\overline{\text{RMSE}}$); B.probability distributions(mean Kullback–Leibler divergence, $\overline{\text{KL}}$), C. phase-space dynamics(relative correlation dimension, $R_{D_2}$). Surrogate 1 achieves lower $\overline{\text{RMSE}}$, while Surrogate 2 performs better in probabilistic and topological metrics. D–H| Neuroscientific perspective. Metrics include D. spatial–temporal similarity, E. spectral similarity, F. FC similarity, G. behavioral decoding accuracy, and H. clinical efficacy in predicting lesion outcomes. Panels D and E focus on the surrogate brain's ability to replicate biological signal characteristics, Panels G and H evaluate its capability to support downstream tasks. Yellow traces denote real brain data; black traces represent surrogate brain outputs.
• Figure 5: Applications of surrogate brain. A. System analysis of surrogate brain. The surrogate allows direct inspection of fundamental system properties, including phase portrait exploration, neural manifold geometry, and network dynamics profiling. B. Surrogate as simulation platform. As an in-silico testbed it enables neural dynamics prediction, virtual surgery (e.g., locating epileptogenic foci to aid clinical planning), and counterfactual experiments such as perturbing connection weights or the excitation–inhibition ratio to probe their functional roles. C. Surrogate guiding neural stimulation. For both non-invasive (e.g., TMS, tES) and invasive (e.g., DBS, SEEG) stimulations, the surrogate brain helps determine where to stimulate, when to target, and how to optimize stimulation parameters.