IRT SystemX’s Post

Following on from MolFlux, we’re also releasing PhysicsML today: an innovative package for physics-based/related models, providing a standardised and modular interface for handling, building and training 3d models with an emphasis on neural network potentials (NNPs). By providing plugins to popular molecular dynamics engines like OpenMM and ASE, PhysicsML facilitates the journey from a trained model to running simulations with NNPs. PhysicsML is freely available and ships with prominent state-of-the-art architectures (including ANI, MACE, Allegro) and can be easily extended to feature new architectures. We also provide access to commonly used quantum mechanical datasets (such as ANI1x, ANI2x, and SPICE) for model training. You can find all the code and documentation here: https://lnkd.in/eAC-sWcf and we look forward to the community contributing to expand the PhysicsML catalogue. #neuralnetpotentials #quantummechanics #3d

Hallucinations don’t just plague LLM adoption. The incorrect prediction of machine learning models also plague efforts to leverage machine learning in science. Hallucinations are particularly insidious in physics based simulations because small errors compound into large errors and infrequent but large errors derail simulations. At Physics Inverted Materials (PHIN), our breakthrough uncertainty quantification allows us to eliminate hallucinations in our machine learning models making them indistinguishable from density functional theory (DFT). PHIN Materials believes that only by creating trustable machine learning models can we leverage the significant performance benefits of ML, paving the way for #DigitizingMaterialsDevelopment. Read more about our state of the art technology, PHIN-atmoic, at our latest blog post ➡ https://lnkd.in/eFzJM-wP #machinelearning #materialsdevelopment #RandD #ai4science #PHINMaterials

🏁 Revving Up Innovation: GM Motorsports Takes the Lead with AI Physics on Rescale! 🏁 Exciting times ahead for motorsports enthusiasts! GM Motorsports is redefining the game with groundbreaking AI-accelerated aerodynamics optimization. 🚗💨 In the high-stakes world of Formula One racing, where every millisecond counts, the Andretti Cadillac team is leveraging AI Physics on Rescale, powered by NVIDIA, to fine-tune chassis aerodynamics like never before. By harnessing the latest technologies from NVIDIA, Microsoft Azure, and Navasto, GM is not just speeding up cars on the track but also accelerating R&D decision-making with unprecedented efficiency and performance. 🌐🔬 Interested in learning how this innovative solution can accelerate your organization's R&D? Reach out or DM me to schedule a demo or a 1-on-1 meeting to explore the possibilities! #GMMotorsports #AIInnovation #Rescale #FormulaOne #NVIDIA #MicrosoftAzure #Navasto #Aerodynamics #EngineeringExcellence

▶️ See it in action: AI Physics Powered by NVIDIA on Rescale - Explore the first turnkey, full-stack AI-accelerated R&D platform! In the dynamic field of computational science and engineering, AI plays a pivotal role by automating critical tasks like data collection, model development, optimization, and design exploration. The integration of neural networks is particularly exciting, offering efficient approximations for complex simulations through streamlined algorithms and GPU acceleration. Rescale’s AI Physics platform, powered by NVIDIA, seamlessly merges applied AI Physics with cutting-edge NVIDIA technologies. From weather predictions to Molecular Dynamics, Biomechanics simulations, Cardiovascular CFD, Heat Exchanger Optimization, and Turbomachinery flows. In this demo video, see how users gain access to optimized infrastructures via a straightforward workflow submission process. #AIPhysics #AI #NVIDIA #Engineering

Physics simulators written from first-principles are not perfect, but don't throw out the baby with the bath water! You're almost never going to develop or code 100% accurate models of the real world for new types of phenomena, and when you do, you're going to spend a ton of time to do so. On the other hand, if you use ML to learn physics model-free, you're throwing out a lot of useful information, and you're going to need a ton of data. We show that even in partially-observable physical systems such as soft robots, relatively small datasets can augment base simulators with residual physics and provide superior accuracy in dynamical settings, and even faster simulation runtimes. We're excited to push this approach to more challenging problems in soft robot modeling in the near future! https://lnkd.in/gsjajMcy Joint work with Junpeng Gao, Mike Yan Michelis, and Robert Katzschmann.

Real-time Physics simulations and Machine Learning convergence is the next big thing in industrial AI and the productive metaverse. Working on it at MYWAI IM me if you have any interest on the topic. https://lnkd.in/djkZeEnX

Learn about BITS: Bi-Level Imitation for Traffic Simulation from the #NVIDIA Research team. BITS is a bi-level imitation learning model that captures the complexity of real world traffic with incredibly fidelity while outperforming previous modeling methods. BITS improved coverage and diversity in traffic scenarios over the next best-performing model by 64% and 118%, respectively, and lowered failure rates by 36%, in a trial conducted for the paper.

PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. We adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. Stanford University Carnegie Mellon University Technical University of Munich and Google paper: https://lnkd.in/ejcPUrum model: https://lnkd.in/etJMMwcV code: https://lnkd.in/etJMMwcV

🎉 Exciting News! 🎉 Just upgraded my toolkit with an RTX3090 GPU! 💻 If you or someone in your circle needs virtual screening assays, molecular docking, or MD simulations📊 to complement your experimental findings, I'm here to help!🤝 🚀 Beyond that, if you share a passion for bioinformatics and are ready to embark on impactful research initiatives, I invite you to connect with me! Let's collaborate and drive innovation together. 🌟💡 #Bioinformatics #Research #Innovation #Science #Collaboration #VirtualScreening #MolecularDocking #MDSimulation #RTX3090 #GPU #Networking

Great read! I especially loved this quote: "While I bow down to our supreme AI overlords in pursuit of this future, we are not there yet." It seems like the file format really matters here, correct? Most of the text-to-3D frameworks I've seen are generating meshes or STL/OBJs. Would it not make more sense to use a 3D data structure that's more workable with an LLM? Like a JSON or YAML file of a CSG tree? Correct me if I'm wrong here. I assume they are using meshes due to the abundance of data to train on. Additionally, it feels like vision capabilities are a must for the future of this workflow. Whether it's used to define a part's bounding area, constraints, or other specific requirements - it would be much easier and faster for an engineer to provide those reqs via an image, rendering, or diagram rather than via text description. You did link to an interesting paper about VLMs for engineering design -- I still need to read it. Either way, very interesting write up. And it's really cool how you used this tool in an unintended way to explore possibilities for engineering. I'm super excited to see where this is heading.