- 14
- 20 339
Open Catalyst Project
เข้าร่วมเมื่อ 4 พ.ย. 2021
Using AI to model and discover new catalysts to address the energy challenges posed by climate change.
Datasets, evaluation and challenges | Open Catalyst Intro Series | Ep. 7
Episode 7: In this episode, we explore how well the ML models we’ve discussed perform in practice. We describe the Open Catalyst datasets and how models are evaluated on them. We conclude by discussing several of the remaining open problems in this space.
This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore machine learning problems in chemistry and material science.
#opencatalyst #ai4science #climatechange
Additional materials:
Open Catalyst Github: fair-chem.github.io/
Open Catalyst Leaderboard: opencatalystproject.org/leaderboard.html
Open Catalyst 2020 Dataset: fair-chem.github.io/core/datasets/oc20.html
OpenDAC23 Dataset: fair-chem.github.io/core/datasets/odac.html
Model Checkpoints: fair-chem.github.io/core/model_checkpoints.html
This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore machine learning problems in chemistry and material science.
#opencatalyst #ai4science #climatechange
Additional materials:
Open Catalyst Github: fair-chem.github.io/
Open Catalyst Leaderboard: opencatalystproject.org/leaderboard.html
Open Catalyst 2020 Dataset: fair-chem.github.io/core/datasets/oc20.html
OpenDAC23 Dataset: fair-chem.github.io/core/datasets/odac.html
Model Checkpoints: fair-chem.github.io/core/model_checkpoints.html
มุมมอง: 830
วีดีโอ
Equivariant Models | Open Catalyst Intro Series | Ep. 6
มุมมอง 6K4 หลายเดือนก่อน
Episode 6: In this episode, we explore ML models that have equivariant representations. These model representations are quite fascinating, since they change predictably given changes in the input. For instance, if the input atoms are rotated, the model’s internal representation will also “rotate”. We’ll discuss how a special set of basis functions called spherical harmonics are used in equivari...
Invariant Models | Open Catalyst Intro Series | Ep. 5
มุมมอง 8334 หลายเดือนก่อน
Episode 5: In this episode, we dive into ML models that have invariant representations and how forces are estimated. A physical property of atomic systems is their energy should be invariant to system rotations. Invariant models have internal representations that do not change when a system of atoms is rotated, which ensures rotation invariance of the energy prediction. This video series is aim...
ML Basics for Modeling Atoms | Open Catalyst Intro Series | Ep. 4
มุมมอง 1.5K4 หลายเดือนก่อน
Episode 4: How do we model atoms using machine learning and AI? In this episode, we cover the basics of how systems of atoms are modeled using Graph Neural Networks. This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore machine learning problems in chemistry and material science. #opencatalyst #ai4science #climatechange ...
How do we model catalysts? | Open Catalyst Intro Series | Ep. 3
มุมมอง 1.4K9 หลายเดือนก่อน
Why are catalysts important, what are they, and how do we model them computationally? We’ll answer all those questions in this video and how discovering new catalysts could have a meaningful impact on climate change. This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore machine learning problems in chemistry and material...
How do we model atoms? | Open Catalyst Intro Series | Ep. 2
มุมมอง 1.7K10 หลายเดือนก่อน
How do we model atoms using Density Functional Theory (DFT) and why is this so computationally expensive? This video gives a short overview of traditional approaches to modeling atoms, and why machine learning is needed to speed up these calculations. This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore machine learning...
Why model atoms? | Open Catalyst Intro Series | Ep. 1
มุมมอง 2.2K10 หลายเดือนก่อน
To get engaged with the Open Catalyst project check out opencatalystproject.org/. Episode 1: Why is the problem of modeling atoms so important? We discuss just a few of the many examples of how better models of atoms and chemistry could have world changing impact. This video series is aimed at machine learning and AI researchers interested in gaining a better understanding of how to explore mac...
NeurIPS 2023: Open Catalyst Challenge | Challenge Overview
มุมมอง 1.2Kปีที่แล้ว
Members of the Open Catalyst Project discuss this year's NeurIPS 2023 Open Catalyst Challenge and how you can get started and participate. 00:00 Introduction 00:25 Motivation 05:26 Task 06:14 Evaluation Metric 09:57 Data 14:01 Approaches Full details: opencatalystproject.org/challenge.html. Questions? Feel free to ask on our discussion board! discuss.opencatalystproject.org/
NeurIPS 2022: Open Catalyst Challenge | 2nd place - Atomic Architects, MIT
มุมมอง 589ปีที่แล้ว
Team Atomic Architects from MIT presenting their 2nd place entry, Equiformer SCN, at the NeurIPS 2022 Open Catalyst Challenge. More details: opencatalystproject.org/challenge.html
NeurIPS 2022: Open Catalyst Challenge | 1st place - Tencent AI Lab
มุมมอง 528ปีที่แล้ว
Team TTRC (formerly Team Tencent AI Lab), presenting their 1st place entry, GeoEnsemble, at the NeurIPS 2022 Open Catalyst Challenge. More details: opencatalystproject.org/challenge.html
NeurIPS 2022: Open Catalyst Challenge Event
มุมมอง 767ปีที่แล้ว
NeurIPS 2022 Open Catalyst Challenge Event. Including a challenge overview, runner-up and winner talks, invited talks by Gabor Csanyi and Yoshua Bengio, and an interactive discussion on OCP and the community. 00:22:50 - Challenge Overview 00:36:12 - Invited talk: Gábor Csányi 01:12:10 - Runner-up talk: Equiformer SCN, Atomic Architects MIT 01:28:35 - Winner talk: GeoEnsemble, Tencent AI Lab 01:...
NeurIPS 2021: Open Catalyst Challenge | Discussion Session
มุมมอง 5272 ปีที่แล้ว
NeurIPS 2021: Open Catalyst Challenge discussion session. Have further questions? Visit: discuss.opencatalystproject.org/ 0:00:00 Introduction 0:01:58 2nd Place Presentation 0:14:13 2nd Place Q&A 0:21:05 1st Place Presentation 0:31:12 1st Place Q&A 0:49:55 Discussion Session
NeurIPS 2021: Open Catalyst Challenge | 2nd place - Innopolis AI
มุมมอง 5462 ปีที่แล้ว
Innopolis AI presenting their 2nd place entry at the NeurIPS 2021: Open Catalyst Challenge. More details: opencatalystproject.org/challenge.html
NeurIPS 2021: Open Catalyst Challenge | 1st place - Microsoft Research Asia
มุมมอง 2.1K2 ปีที่แล้ว
Microsoft Research Asia (previously "MachineLearning") presenting their 1st place entry, Graphormer, at the NeurIPS 2021: Open Catalyst Challenge. More details: opencatalystproject.org/challenge.html
I need to download the dataset. How can I do that?
At time stamp = 5:55 In the time complexity analysis n = number of atoms specified in the system right? Also O(n^2) is for the training time or prediction time? Cause if the prediction time going O(n^2) or higher worse, isn't it better to use the physics-based approaches instead of MLPs?
Brilliant explanation! My only regret is, that I haven't found this earlier.
Suggesting to make the explanations more to the point, with significantly less details and options. Once we're on board, you can point to the more advanced ideas, observations, etc. Writing this after trying for the 3-4 time to follow. I have not given up yet. However, I do want to bring to the attention of the organizers. This is very interesting and important. I do appreciate the project.
Thanks
Awesome video. May I ask how might I apply the Wigner D matrix? I have the sum wave function. Then right now, I am trying to shift it by A that's just a scalar/number and not a matrix. It works for one wave function, but for two it's not moving in unison and it seems to be moving independantly. (Not an issue, as expected). But, now, how do I go about getting the Wigner D and how to apply it? I figured I can do J=1/2, but it feels like I need to configure it with alpha,beta,gamma but what would those values be in this context?
Figured it out! Wigner D, should be initialized with J=0.5, alpha=0, gamma=0, Beta should be k*a, where k is the frequency of the basis function. Again, thanks for an awesome video.
Super content, thanks a lot!
This is such a good introduction to DFT. Thanks!
Too good!
Hi! Really appreciate the effort invested in the videos and the whole project. Rarely one sees true applications of AI for good. I have a couple of questions: 1. I suppose that the ground truth for energy is computed via more computationally complex methods, like DFT. That seems to be the case from the datasets released. How does that work for the larger systems where DFT is computationally impractical? Do we always break down problems to manageable pieces (like the tiling approach described in the video) or do we hope that the model trained on smaller systems will generalize to bigger ones? 2. I suppose that the "gradient approach", where the forces are computed via gradients of the energy wrt. the positions, would give the true (or sufficiently good) forces only if the model is true (or sufficiently good). Is my intuition correct? 3. I suppose that in the "direct approach" the ground truth forces also come from the DFT. Aren't the forces coming out of this model not only easier to compute, but also better (compared to the "gradient approach"), in the sense that they are modelling the ground truth forces (assuming that the DFT is better model than the ML model). 4. You mentioned that the forces should sum to 0 (if I understood well that as a consequence of energy conservation principle, might have assumed too much here), but in direct approach there are no guarantees that this will be the case. I suppose one could add a term to penalize the deviation from zero sum. Is that done usually, or that constraints the model too much, so we end with less useful models this way? (It might be that I missed something).
Hello, I am not the author of the video, but I think I can answer your questions with an educated guess (I am doing my phd on dft and AI). 1) Yes, the bigger the system gets, the more costly it is to compute. It is not rare for groups to ask for computing time in supercomputers to carry out simulations. The purpose of fitting dft with (machine learning) potentials is both to be able to try a lot of systems and to generalise to bigger ones. For example, fitting your model on small molecules like benzenes and trying to predict on large polyaromatic hydrocarbons and molecular networks.
2) this is an opinion one but in principle the gradient computation is better because, by computing it from the gradient, the resulting force is conservative by construction. You don't get that by the direct approach. If you are worrying about not using the forces in training, you can incorporate them in the loss function as well in the gradient approach, but instead of only doing backprop on the loss, you are now doing backprop, first on the energy and then on the loss, 2 backprops for each batch.
4) in principle the forces in dft should sum to 0 because otherwise you are accelerating your whole system, which doesn't make sense. Yes, also in absence of external forces the total momentum is conserved. In reality sometimes these dft code fail a little hehee, but it rarely matters with respect to the degree of accuracy you need for the calculation. If the dft forces Sum to 0 and you are fitting to the dft, you are implicitly adding the constraint to the model. This is an advantage you get when using the gradient approach since the gradient of a minima is 0 ;) You could always add the constraint explicitly in the loss function and it wouldn't be difficult at all. The loss would be new_loss = older_loss + \lambda * (force_prediction.sum()) **2, but you would have to look at how the model implements the prediction. This is just a pseudo code for you.
Thanks a lot for taking the time to reply in such detail! Much appreciated!
@@josipKraap your welcome. It is cool that more people are interested in this. There is life in the deep learning industry beyond ad optimization and LLM ;)
I love these brave women! Thanks for standing up!
Thank you! I'm looking forward to the next video that will cover equivariance. It's a concept I'm trying to master now in my studies :)
Thank you for your great lectures
Does anyone got any sources to a notebook tutorial with a dataset to practice GNNs? I have not really found much.
Thank you for the video. I was trying to run the tutorial notebook on Google Colab but I am unable to install ocp models. Can you please help?
@larryzitnick6444 Larry, I am upset that open catalyst list of adsorbates has almost every single organic molecule except for lithium sulfur polysulfides (Li2Sx). I model them for next generation lithium sulfur batteries that need metal coatings for the carbon cathode. The chemical phenomena is surface catalysis, this seperates them from lithium-ion. Where do I complain to Meta?
Most importantly, when do we get the next video :)?
On the cation walk.
Good explanation, are you guys working with Atomic Tessellator? This video reminds me of their catalyst workflow.
@4:36 why not work with a smaller number of atoms? Are there tools and simulators?
In the simple platinum case, I agree it looks like you could use fewer atoms. However, if you want to understand what surfaces are likely to be realized in practice, you need to model the nano particles with a “reasonable” number of atoms, unless certain surfaces won’t be seen. If you start thinking about modeling chemical reactions or proteins, then the number of atoms you want to model can get really really big :)
Hey Larry, this is a pretty interesting introduction to a what seems to be a tough problem! When can we expect more on this? As someone working in DL/ML but with no chemistry background, I'm looking forward to more
Glad you like them! I’m hoping to finish the next video on catalysts in early December. Not sure if I’ll have time to finish the two ML videos before the new year, but should be out in January.
I have feelings this guy doesn’t know basics of chemistry and physics. So many errors difficult to mention all of them.
Happy to answer any questions or concerns about specific points.
Looks like you guys know position of atom and its speed in your simulation. Or I miss something? Nobel prize, obviously!
In 1998, the inventors of DFT did actually win the Noble prize! All we’re trying to do is approximate the same calculations with machine learning to make them faster.
@@larryzitnick6444I see, you want to apply CS to HPC to solve equations of mathematical physics. The problem is not any CS model has a physical meaning and can cooperate with chemistry principles. I’d like to recommend as a prerequisite to apply CS to physics take at least 1 year class in Statistical Mechanics ( class about statistics of molecular physics, moving and interacting between molecules), nuclear physics (to understand differences between weak and strong forces, differentiation between particles and molecules), organic and nonorganic chemistry. I mean to get a Master of Science in physics first to understand how CS could be applied to model chemical physics phenomena. As a creator video on such complex topic on the edge between physics and chemistry you shall understand meaning of every word and sentence before thinking how that could be interpreted in CS. IMHO, you are an expert in CS and you would like to apply modern advances in your field to new forms of your area of chemical physics. Good intention but requires really good knowledge at chemistry and physics. Nothing wrong, solid knowledge in physics and chemistry will help you to understand how CS maybe used to interpret real physical world correctly. BTW, I’m open for DM. In my past I used to develop computer models of chemical physics phenomena specifically my work was used for simulation of nuclear blast ignition and boosting. That is about critical mass creation by conventional explosives. I highly recommend to edit this video. Good luck.
Kudos!!! Quite inspirational. It takes keen interest and lots of energy to go through so much dense content around chemistry. I mean you did it 20 years after your chemistry class. 👏
Really exciting about merging ML with chemistry
Promo`SM
Who is here at the beginning? 👇
Who is here at the beginning? 👇
I wonder why the comment section is so breezy. There is a guy that post nonsense and gets 400 min. comments and upvotes. I guess the TH-cam teams has gotten to be, for a long time, cynical about the population.