Conclusion
In conclusion, our results illustrate that existing performance metrics such as HV do not really reflect the goal of fleshing out the PF region, where HV-based methods like qNEHVI may not achieve satisfactorily. This reflects an aspect of optimisation which might be neglected in the purview of multi-objective materials discovery: which is to find a diverse set of optimal solutions that can adequately convey the trade-offs between conflicting objectives. We thus present alternative illustrative means such as probability density maps to better benchmark the performance of optimisation strategies for such purposes. Moving ahead, we hope that this can spur further improvement for MOBOs as well as a stronger consideration for the use of MOEAs for materials problems due to its heuristic nature in exploiting the PF.
Acknowledgements
K.H. acknowledges funding from the Accelerated Materials Development for Manufacturing Program at A*STAR via the AME Programmatic Fund by the Agency for Science, Technology and Research under Grant No. A1898b0043. K.H. also acknowledges funding from the NRF Fellowship NRF-NRFF13-2021-0011.
K.Y.A.L. and K.H. conceived of the research. K.Y.A.L. working with E.V-G. and Y-F.L. developed and tested the algorithms and datasets, with key intellectual contributions from all authors. K.Y.A.L. wrote the manuscript, with input from all co-authors.