Félix Therrien
@felix-therrien.bsky.social
Scientist at Mila - Discovering new materials to solve climate change using physics and ML
~~~
Scientifique à Mila - À la découverte de nouveaux matériaux pour résoudre les changements climatiques en utilisant la physique et l'AA
~~~
Scientifique à Mila - À la découverte de nouveaux matériaux pour résoudre les changements climatiques en utilisant la physique et l'AA
I would like to formalize that with a more in-depth study, but I think it aligns with some of the litterature on generalizability:
doi.org/10.1038/s415...
doi.org/10.1038/s415...
A critical examination of robustness and generalizability of machine learning prediction of materials properties - npj Computational Materials
npj Computational Materials - A critical examination of robustness and generalizability of machine learning prediction of materials properties
doi.org
May 9, 2025 at 1:12 AM
I would like to formalize that with a more in-depth study, but I think it aligns with some of the litterature on generalizability:
doi.org/10.1038/s415...
doi.org/10.1038/s415...
We found that ML proxies perform significantly worse on generated molecules than on dataset molecules. A large SOTA model who did great on benchmarks was more unreliable than a small GNN on these molecules, regardless of how they were generated.
May 9, 2025 at 1:12 AM
We found that ML proxies perform significantly worse on generated molecules than on dataset molecules. A large SOTA model who did great on benchmarks was more unreliable than a small GNN on these molecules, regardless of how they were generated.
Personally, one of the main takeaways I got from writing this paper is that *evaluating the performance of generative models is really hard because their performance is super dependent on the ML proxy you use for evaluation.*
May 9, 2025 at 1:12 AM
Personally, one of the main takeaways I got from writing this paper is that *evaluating the performance of generative models is really hard because their performance is super dependent on the ML proxy you use for evaluation.*
The paper shows how you can use molecule property predictors as generators through a simple input optimization if you carefully restrict the input space. It performs surprisingly well and generates super diverse molecules.
May 9, 2025 at 1:12 AM
The paper shows how you can use molecule property predictors as generators through a simple input optimization if you carefully restrict the input space. It performs surprisingly well and generates super diverse molecules.