Deep Adaptive Model-Based Design of Experiments

Researchers Arno Strouwen and Sebastian Micluţa-Câmpeanu have introduced a breakthrough method called Deep Adaptive Design (DAD) that revolutionizes how scientists conduct experiments with nonlinear dynamical systems. Traditional Model-Based Design of Experiments (MBDOE) requires computationally expensive posterior inference and design optimization between each experimental step, making real-time applications impossible. DAD solves this by amortizing sequential design into a neural network policy trained offline, then combining it with differentiable mechanistic models. This approach eliminates the need for costly optimization during experiments while maintaining scientific rigor.

The team developed a transformer-based policy architecture specifically designed to handle the temporal structure of dynamical systems and extended sequential contrastive training objectives to manage nuisance parameters. They demonstrated DAD's effectiveness across four increasingly complex systems: a fed-batch bioreactor with Monod kinetics, a Haldane bioreactor with uncertain substrate inhibition, a two-compartment pharmacokinetic model with nuisance clearance parameters, and a DC motor system for real-time deployment. The method enables scientists to run adaptive experiments where the AI dynamically decides the next best experimental conditions based on real-time results, dramatically accelerating parameter estimation and discovery processes in fields from biotechnology to pharmacology.