Rhys Goodall (@RhysGoodall) header
Rhys Goodall (@RhysGoodall) profile picture

Born 362 ppm. https://teapaw.org/t.co/eXRBdGXem5

Joined May 2011
Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)Photo by Rhys Goodall (@RhysGoodall)
The future is electric and any actions that reduce the number of BEVs in the UK is a costly attack on the inevitable future.
1
1
128
I have prematurely said that MBD was saturated more times than I should admit. Perhaps there is still more juice to squeeze but if I was still doing day to day research in this space I think there's more ripe fruit to pick elsewhere. I would like to see the following:
This quoted post is unavailable.
1
1
525
- R2SCAN datasets for training foundation models like MatPES/MP-ALOE as MLIPs get more leverage approximating higher fidelity simulations at fixed inference cost. This also helps with unifying molecules and materials as R2SCAN + dispersion good for mols/mofs/mats
1
187
Rhys Goodall retweeted
> be me > exponential curve supervisor > get to work > check-in on the curve > it’s sigmoid
32
186
6
4,330
123,175
Another step forward in ML-first materials. We’ve seen machine learning revolutionize industry after industry, and today, we release EGIP, showing our ML-first approach that out-performs traditional physics-based methodologies: radical-ai.com/news/EGIP
1
5
21
2,051
TorchSim v0.2 is out! github.com/Radical-AI/torch-… I am especially happy to have 5 new contributors and 3 new model interfaces. Below, a sneak peek of a WIP figure describing TorchSim's approach to memory estimation.
5
1
24
1,730
Rhys Goodall retweeted
Today we release TorchSim, a next-generation PyTorch-native atomistic simulation engine for the MLIP era. Our hope is that this open-source package ignites a materials simulation revolution.
6
71
7
354
50,933
Is there any good literature on the use of deep set models as part of deep kernel learning GPs for bayesopt? @Ryan__Rhys @andrewgwils
1
1
225
In relation to the noble prize in physics I've seen several articles saying that @geoffreyhinton worked on applying ML to improve solar cells, if anyone has the source material or references I would love to read them but haven't found them through search.
1
2
463
Orders of magnitude: coal vs. lithium.
57
325
41
1,463
167,243
Rhys Goodall retweeted
the bay area is one of the greatest failures of urban planning in human history - best weather on earth - waterfront property, islands, and hills for views - massive natural bay what we got: - mindless suburban sprawl - a “subway” used by <1% of the population
106
377
59
6,561
450,047
Rhys Goodall retweeted
If it seems like I’ve been talking about data centers a lot, here’s a good example why. Major climate goals are at risk. “Microsoft’s predicament is one of the first concrete examples of how the pursuit of AI is colliding with efforts to cut emissions.” bloomberg.com/news/articles/…
11
30
5
107
11,850
The philosophy of Matbench Discovery was to try produce a discriminative benchmark that mirrors a real workflow in High-Throughput inorganic crystal discovery. We’re hoping to commodify the thermodynamic screening steps such that researchers can focus on properties and synthesis
It’s great to see that ML forcefields slowly start to be distributed in versions and packages. The field needs this sort of consolidation and adoption to move from experimenting to production.
1
11
2,237
Rhys Goodall retweeted
Using the recently added MACE multi-GPU support, @PhilippBenner2 burned 1440 A100 hours to train a 16M param universal MACE on MPtrj that achieves a new SOTA on Matbench Discovery, reaching F1 score of 0.64 in crystal stability prediction! 🧵
3
11
2
53
17,734
In a manufacturing context we’re often running multiple repeats such that we get an understanding of process variances. When using BO there seems to be a gap in my literature search about how to effectively use repeat noisy measurements. Explicitly hetroskedatic GP models 1/
1
1
279
Suffer from numerical issues but there is much less work exploring demanding BO problems with other classes of surrogates. Are there any suggestions of recent work in this area I may have missed? 2/2
1
144
Rhys Goodall retweeted
still thinking about the skyline of San Jose, the tech capital of the world I was expecting futuristic utopia vibes and instead got, like, Plano Texas vibes
114
112
40
3,288
418,651