Optimizing Biophysical Large-Scale Brain Circuit Models With Deep Neural Networks.

Biophysical modeling provides mechanistic insights into brain function, from single-neuron dynamics to large-scale circuit models bridging macro-scale brain activity with microscale measurements. Biophysical models are governed by biologically meaningful parameters, many of which can be experimentally measured. Some parameters are unknown, and optimizing their values can dramatically improve adherence to experimental data, significantly enhancing biological plausibility. Previous optimization methods - such as exhaustive search, gradient descent, evolutionary strategies and Bayesian optimization - require repeated, computationally expensive numerical integration of biophysical differential equations, limiting scalability to population-level datasets. Here, we introduce DELSSOME (DEep Learning for Surrogate Statistics Optimization in MEan field modeling), a framework that bypasses numerical integration by directly predicting whether model parameters produce realistic brain dynamics. When applied to the widely used feedback inhibition control (FIC) mean field model, DELSSOME achieves a 2000× speedup over Euler integration. By embedding DELSSOME within an evolutionary optimization strategy, trained models generalize to new datasets without additional tuning, enabling a 50× speedup in FIC model estimation while preserving neurobiological insights. The massive acceleration facilitates large-scale mechanistic modeling in population-level neuroscience, unlocking new opportunities for understanding brain function.