Janosh
@jrib_
Followers
679
Following
2K
Statuses
727
Working on data-driven autonomous materials discovery at Radical AI, prev @ Materials Project (LBNL).
NYC
Joined October 2020
re 3: true for now but fine-tuning MLIPs directly on experimental measurements in an attempt to surpass DFT accuracy at least for certain property predictions is the obvious next step. the easiest way is probably to add Hessians to the loss function and back-propagate on phonon-derived observables such as thermal conductivity. the tricky part might be to not sacrifice too much energy/force prediction accuracy in the process if you still care about that
Nice tool, glad they open-sourced it, but: 1. I believe @Robert_Palgrave has opined, critically, on this statement in previous comments: "... MatterGen, with the caveat of compositional disorder between Ta and Cr." I think this is problematic, rather than a minor point. It suggests that ground-state structure prediction can be misleading when described as "discovery." In reality, finite-temperature effects and kinetic factors play significant roles, which means such predictions do not fully capture the necessary temperature & processing complexity (obviously). 2. In my opinion, just predicting a "novel" composition and structure that is GS stable with ideal properties is insufficient for practical use by materials synthesis chemists. We must provide guidance on a "recipe" for synthesis. Or, more future focused, integrate this approach into a self-driving lab with both a prediction head (structure/property prediction) and a process head (synthesis/process optimization) would make it more actionable and improve prediction head via RLHF. Others are doing this, but my hope is they are doing so in a vertical integration schema (@josephfkrause ?) . 3. It is essential to recognize that, to my knowledge, ML/AI models used for material predictions are strictly trained on DFT data, meaning their predictions are inherently constrained by the accuracy and limitations of DFT. It is an approximation to the many-body e-e interactions by reformulating the problem into an effective single-electron system, corrected by our best informed guess for an electron density functional. While DFT works well for many material predictions, its approximations falter when dealing with systems where strong e-e correlations dominate (e.g., superconductors, Mott insulators). Therefore, caution and skepticism are necessary when applying DFT-based ML/AI models to such materials. Feel free to criticize my comments, this is meant to engage.
0
0
14
@XirtamEsrevni @TimothyDuignan @cyrusyc_tw very nice project! haven't used it yet either but lots of great visualizations. MLIPX makes it very easy to compare different models to understand their tradeoffs and limitations which is urgently needed. really like the NEB example
1
0
3
RT @SimoncelliMic: Our k_SRME benchmark for foundational ML potentials (fMLPs)—used on the Matbench platform to test fMLPs' accuracy in des…
0
2
0
RT @wellingmax: “Global warming is on the cusp of crucial 1.5 °C threshold, suggest ice-core data.” Face the facts: this…
0
5
0
RT @getjonwithit: Put your work on arXiv and release 100% of your code/data/Mathematica notebooks/etc., packaged up in such a way that *any…
0
378
0
@andrewwhite01 I agree it would be nice if more people who rightly complain about lack of data would rise to the challenge and start generating some.
1
0
5
@XirtamEsrevni Credit to Balazs Pota, @AhlawatParamvir and Michele Simoncelli who did most of the work!
0
0
2
@venkatkapil24 @lonepair Entirely energy based which was a significant shortcoming of Matbench discovery up until now. Forces only indirectly contributed to better metrics by enabling a more informed energy prediction if a model accurately recovers the DFT ground state during relaxation
1
0
3
RT @AhlawatParamvir: Introducing SRME: a new metric for further assessment of foundational machine learning interatomic potentials! 👉A sm…
0
13
0
RT @taras_y_sereda: A wonderful collection of TikZ figures. Thank you @jrib_ for maintaining this repo https://t.co…
0
1
0
actions speak louder than words. open science and open data is often quoted but not always practiced. really appreciate the fairchem team (@lbluque , @mshuaibii , @xiangfu_ml , @bwood_m , @csmisko , Meng Gao, @ammarhrizvi , C. Lawrence Zitnick, @zackulissi) leading by example here and sharing everything freely: data, weights and code! shows how much academia and industry can complement each other when academia lays theoretical foundations that point the way and then industry can put major compute behind those methods to really scale them up and leverage computational tools to their full potential
Introducing Meta’s Open Materials 2024 (OMat24) Dataset and Models! All under permissive open licenses for commercial and non-commercial use! Paper: Dataset: Models: 🧵1/x
2
6
63
@Andrew_S_Rosen @Sergei_Imaging @olexandr I start really trusting experimental results once they were reproduced on different equipment and ideally by a second party
0
0
1
RT @MarkNeumannnn: Orb-v2 also extends the Matbench Discovery leaderboard to 88F1, improving upon Orb-v1 by 2F1. We'll be releasing a techn…
0
1
0
Agree with both: I think Cursor and 3.5 Sonnet are both underrated. Definitely a killer combo!
interesting that so many people feel like this - I feel the opposite! I like Cursor because of: 1. Line level diffs 2. tab to next semantic location 3. "apply to file" from a chat 4. All the ways you can reference things in the chat 5. The models It's UX all the way down
0
0
2