Zhaocheng Li

PhD Student in the Department of Mathematical Sciences

What do you do?

I am currently a PhD student in the direction of models surrogate and uncertainty quantification for computationally intensive models in the Department of Mathematical Sciences at Durham University.

How are you involved in this area of science? 

I am constructing the emulator based on Bayes Linear Analysis to calibrate the simulator that models the real-world system. Many simulators are too computationally intensive and expensive to iterate till the acceptable choice of parameters or inputs appears.

Our methodology aims to emulate the simulator to accurately predict its output while providing the uncertainty quantification for the result by letting the simulator run several inputs/parameters to come up with some certain results. We apply the priori beliefs from the stationary process e.g. Gaussian Process to the computation of the exponentiated correlation among each input to derive the posteriori in terms of adjusted expectation and adjusted variance with respect to some certain running inputs. This largely reduces the cost of the model running. I am struggling to build up a surrogate for the model that contains locations of multiple partial discontinuities (e.g. oil reservoir prediction in a place with geological faults).

What do you love about this topic?

This direction is one of the novel branches of the whole AI field with huge potential application perspectives. There is an increasing trend in the construction of models by applying huge amounts of data, which always leads to the expensive model running with tremendous cost of energy. Our methodology uses several running input data but will predict the global behaviour of the given data, and we already successfully emulated models of galaxy formation, oil reservoirs, plant hormones, etc. However, as a new technique, much more complex models like deep-layer Neural Networks lack suitable emulations. The unknown is always exciting!

How does this work deliver real-world impact?

Computational modelling helps scientists better understand and predict real-world systems, but it also relies too much on data and parameter size for making a credible analysis. Our work has already shown the possibility of using several certain running inputs to predict the whole complex model's output for some complex real systems. Hence, we will promote the performance of computational modelling from the following perspective:

• Save energy.
• Save computation cost.
• Provide a new heuristic understanding of alternative AI construction.
• Prove the possibility of predicting the behaviour of the infinite output space by the finite input size.