Intelligent Data Collection

This covers a broad set of methods and applications in which we wish to optimise an experiment, or a sequence of experiments, to achieve a specific goal. For example, (1) be maximally informative about some factor of interest (traditional design of experiments); (2) do model selection or (3) maximise some other quantity of interest (Bayesian optimization). We have championed ground-breaking advancements in adaptive and efficient design space sampling, crucial for optimizing specific objectives, notably entropy-based ones.

While these methods have clear implications for manufacturing and synthetic biology, their profound impact extends to fields like material design and drug development. We have introduced pioneering methods in optimal sequential design of experiments, as documented in our work presented at ICML 2022 (Blau et al., 2022). Furthermore, we have offered a transformative take on Bayesian optimization, recasting the problem as a probabilistic classification task (Tiao et al., 2021). This innovative viewpoint enables the application of modern probabilistic classifiers, such as deep neural networks, convolutional neural networks, and transformer-based models. The outcome is a leap in model expressiveness, versatility, and scalability.

Recent Developments

Bayesian Optimization and Calibration of Computer Models

• Bayesian optimization as classification: BORE (Tiao et al., 2021)
• Bayesian adaptive calibration and optimal design (Oliveira et al., 2024)
• Generative Bayesian optimization: generative models as acquisition functions (Oliveira et al., 2026)
• Thompson sampling in function spaces via neural operators (Oliveira et al., 2025)
• Variational search distributions (Steinberg et al., 2025)
• Amortized active generation of Pareto sets (Steinberg et al., 2025)

Optimal Desing of Experiments (DoE)

• Adaptive DoE via reinforcement learning (Blau et al., 2022)
• Cross-entropy estimators for adaptive DoE (Blau et al., 2023)