Scientific Machine Learning

Motivation

Physics-based simulations can resolve electric fields, radiation transport and transient charge collection in great detail, but a large design sweep may require many expensive runs. Scientific machine learning can learn relationships between design variables and simulation outputs, allowing promising regions to be explored more efficiently.

This is a developing research direction within The Open Notebook. The aim is to complement physical simulation—not replace it.

From simulation data to surrogate models

A workflow begins with a carefully designed set of TCAD or particle-transport simulations. Inputs may include geometry, materials, doping, bias and radiation conditions. Outputs may include field metrics, breakdown indicators, deposited energy, peak current, collected charge, collection time and response uniformity.

A surrogate model is trained on these results and evaluated on unseen simulations. Once validated, it can estimate responses quickly enough for sensitivity studies, optimization and uncertainty analysis.

Research questions

1. Which design variables control detector response most strongly?
2. How many simulations are needed before a surrogate becomes reliable?
3. How should physical constraints and known scaling laws be included?
4. How can uncertainty be reported so that a fast prediction is not mistaken for a verified device result?
5. Can active learning select the next simulation that is most informative?

Methods under exploration

• Regression baselines and response-surface models
• Gaussian-process surrogates with uncertainty estimates
• Tree-based ensemble methods
• Neural-network models for higher-dimensional data
• Sensitivity analysis and feature importance
• Bayesian optimization and active learning
• Physics-informed features and constrained predictions

Principles

• Physics first: models should respect governing physics, symmetry and limiting behaviour
• Validation beyond a random split: test new regions and physically meaningful edge cases
• Uncertainty is part of the result: show where new high-fidelity simulation is required
• Reproducible workflows: use structured datasets, versioned notebooks and clear metadata

Intended outcomes

• Faster screening of detector designs
• Identification of dominant physical parameters
• More efficient allocation of high-fidelity simulations
• Compact models for design exploration
• Open teaching examples connecting semiconductor physics, statistics and machine learning