Simulation has become a cornerstone of modern-day research, relied upon in practically every quantitative discipline. Simulations are used to model stars, create digital twins, predict the weather, our economy, clinical trials, pandemics, and even the movement of individual cells. Given the ubiquity of simulation in science, it is vital that researchers have up-to-date knowledge of robust methods for developing and applying simulations. Model assessment, especially in the context of parameter estimation, can be achieved using Bayesian inference approaches. Given a model and set of observations, Bayesian inference evaluates (un)certainty on model parameters. Importantly, recent advances in Bayesian methods, alongside computational speed, and deep learning methods, allow for fast and efficient parameter estimation and model selection with few datapoints. This interest group brings together early career researchers with world-leading scientists to discuss simulation-based scientific methods. This interest group provides a platform for scientists at the Turing and across the UK to engage in conversations on advancing simulation-based science.
Explaining the science
Simulation-based inference uses deep learning methods such as normalizing flows to approximate the posterior or likelihood for faster and more efficient parameter estimation. These methods can be extended to account for model misspecification or combined with embedding models to parameterise models using inferred model features.
Many simulations are prohibitively expensive to generate sufficient methods for sampling-based approaches such as simulation-based inference or approximate Bayesian computation. This is particularly true in fields such as climate and earth sciences. Surrogate models address this by introducing smaller models which learn a mapping between model parameters and initial conditions and output, considerably speeding up model evaluation.
Typically, model calibration and parameter estimation are done using experimental observations. However, many disciplines such as in life sciences, chemistry, and robotics allow for experimentation to test model assumptions and output. Active learning attempts to identify experiments that give the greatest information about model validity. These methods are often applied in drug discovery to identify candidates for clinical trials.
Fundamental to simulation-based methods is a generative model that is able to reproduce salient features of the system under investigation. These models have traditionally been developed through mathematical equations such as ordinary and partial differential equations. Hybrid models augment these mechanistic models with deep learning components to retain explainability but improve model power.
This interest group provides a knowledge exchange with a focus on the most recent advances in simulation-based inference and their applications in multiple disciplines. The main goal is to promote the exchange of ideas and create a supportive environment to work on similar data analysis challenges encountered in multiple research areas. In particularly, we are interested in:
- Developing methods that tightly couple simulations to observations
- Identifying new approaches for surrogate models of costly simulators
- Highlighting methods for improving model predictions through the application of experimentation and active learning
- Advancing the field of simulation through the combination of mechanistic models with deep learning
How to get involved
- How can we retain the explainability of mechanistic models whilst improving their predictive power?
- How do we separate out epistemic and aleatoric errors in model evaluation?
- How do we most effectively deploy simulations to advance science and how do we identify simulations that can lead to incorrect conclusions?
- How do we improve trust in how simulations are used in society?
- How can we ensure simulations are robust to new evidence?