Emulator trained on 20,000 simulated crystallisations benchmarks drug-process optimisation strategies

The random forest learned from the population balance model's output ^[3], so its under-1% error describes how faithfully it copies that model. The model was in turn calibrated with experimental data from the Scale-Up CMC DataFactory, a self-driving lab ^[2]. A strategy that scores well on the emulator has beaten a simulation two steps removed from a real crystalliser. The forest approximates the model, and the model approximates one generic cooling-antisolvent crystallisation ^[1].

The starting designs mix Latin hypercube sampling, Sobol sequences and random sampling at several sample sizes ^[4]. If each of the twelve was paired with each of the six acquisition functions, there are 72 pairings ^[1], and the 360 campaigns ^[5] would come to five per pairing if the runs were spread evenly ^[2]. Random Sampling and Space Filling sit in the acquisition list beside EHVI, NEHVI and NParEGO ^[5]. The first two ignore the Gaussian process's predictions of where good results lie, so the hypervolume-seeking methods have a control to beat.

The authors also put a less standard optimiser, gradient boosting, through the emulator. A Gaussian process with expected improvement beat it by between 6% and 12% across differing initial conditions ^[7]. The largest margin is twice the smallest ^[4]. How a campaign starts changes how much the choice of optimiser matters.

The thing this doesn't tell you is how a hypervolume gap turns into lab time. Hypervolume ^[6] scores the set of trade-offs a campaign ends with. It does not count the experiments spent getting there, or the real runs that fail and have to be repeated. Running 360 campaigns ^[5] was practical because each one cost only compute.

The article is open access under a non-commercial, no-derivatives Creative Commons licence ^[10]. The abstract does not say whether the emulator itself is released as code, or which design and acquisition pairing came out on top.

I think the emulator is most useful for the job the authors describe: evaluating new and existing strategies against one another under realistic pharmaceutical conditions ^[8]. It is weaker as a forecast of how many real experiments a campaign will need. That use would require the population balance model to hold at every condition an optimiser visits, and the validation figure in the abstract is for the forest measured against the model ^[3].