Dr. Craig Bower
I research how machine learning, mathematical modelling and uncertainty can be combined to build intelligent systems that learn, predict and make decisions when information is limited, expensive or incomplete.
A recurring question across my work is: how can an intelligent system learn reliably when observations are sparse, uncertainty matters and acquiring new information is costly?
Selected research
HyperFlux
A self-learning scientific digital twin combining ecosystem physics, graph neural networks, uncertainty-aware prediction and sequential local adaptation.
Themes: Scientific ML · Digital twins · GNNs · Online adaptation
Variable-Interaction Graph Networks (VIGNet)
Learning coupled physical systems from the interactions that actually exist. VIGNet builds neural-operator connectivity directly from PDE variable dependencies, with provable results for approximation, required depth, generalisation and graph mis-specification.
Themes: Scientific ML · Coupled PDEs · GNNs
ABBDA
Efficient monitoring and change detection in massive data streams when observing everything is prohibitively expensive.
Themes: Bandits · Sequential decisions · Change detection · Partial information
Adaptive Scientific Campaigns
A research programme asking how expensive simulations and experiments should be selected to maximise defensible scientific knowledge per unit of cost.
Themes: Autonomous science · HPC · Uncertainty · Adaptive experimentation
Physics-Enforced Continuous Neural Operator
A continuous neural-field surrogate for AGN jet simulations combining hard physics projection, adaptive conformal uncertainty calibration and controlled generative recovery of smaller spatial scales.
Themes: Scientific ML · Neural fields · Uncertainty quantification · Physics enforcement
Explanatory Computation
Themes: Explainable AI · Symbolic Regression
Bandit Neural Architecture Search
Themes: Scientific Digital Twins · Multi-armed Bandits · Neural Architecture Search
Adversarial Thresholding Semi-Bandits
Themes: Non-stochasticity · Thresholding Bandits · ALICE Transition Radiation Detector Control
Research direction
My work sits at the intersection of scientific machine learning, sequential decision-making, uncertainty quantification and autonomous scientific discovery. The applications vary, but the methodological problem is often the same: we need to make useful predictions or decisions before we have complete information.