Multi-fidelity Bayesian Optimisation (MFBO)

What is MFBO?

MFBO allows scientist to combine and pool data from different experimental sources. For example, there could be a target experiment, such as a production facility as well as some pilot plants. The production plant is expensive to experiment on, while the pilot plant can produce 10 samples a day with a much smaller budget than production. By pooling these two data sources, we can get to the optimal operation parameters faster than with just the production plant. See our resarch for more details.

Why does MFBO work?

By pooling data, we can exploit the correlations between different datasets. In simple terms, the experiments are different, e.g. a pilot plant and a production plant, but at the core they produce the same material. Hence, when we change the temperature for material production, this will impact the material strength in similar ways. 

We exploit these similatiries to steer the optimisation campaginin the most efficient way, learning simualtaenosuly from the low and high fidelity, ultimately saving research budget and reaching the optimum parameters faster, see our case study.