16 March 2021
Los Angeles, CA
Beyond Limits is inventing the future at NVIDIA GTC on April 12-16, 2021.
GTC is where prestigious scientists and industry pioneers share their breakthroughs in AI, data centers, accelerated computing, healthcare, intelligent networking, game development, and more. Join us at this virtual conference for AI innovators, technologies, and creatives.
Beyond Limits CTO of Industrial AI, Shahram Farhadi will be speaking on Leveraging GPU-accelerated Hybrid AI Frameworks for Energy on Monday, April 12.
Hybrid AI frameworks have seen widespread adoption across industries recently. These frameworks allow the integration of legacy simulation engines and domain knowledge with modern deep learning (DL)/machine learning (ML) solutions. The GPU-accelerated training of industrial-scale, high-fidelity DL models allows for accurate predictions while honoring constraints from related physical equations; domain knowledge offers transparency and explainability.
Shahram’s talk gives an overview of hybrid modeling paradigms using an upstream example of a novel GPU-centric field planning solution. This solves the problem of determining drilling locations, given reservoir geological characteristics, to increase efficiency and value while reducing environmental impact. Hybrid solutions combine simulation-based reinforcement learning with a graph-based, knowledge-driven advisor. Shahram will discuss hybrid modeling benefits and show runtime benchmarks on A100 GPUs – then end with an outlook for hybrid AI and HPC-based workflows across energy, including power and utilities.
Beyond Limits Senior Machine Learning Scientist, Vidyasagar will be presenting a technical session on GPU-accelerated Deep Reinforcement Learning for Field Planning on Tuesday, April 13.
Recently, GPU-accelerated reinforcement learning (RL), using deep neural networks, has gained significant traction. Vidyasagar will describe RL applied to the sequential decision-making of well placement in the upstream energy industry. High-fidelity simulations of geological subsurface reservoirs are coupled with the placement of sources/sinks corresponding to wells; the goal is to maximize net present dollar value (NPV) obtained through cumulative production. Due to high-resolution grids and combinatorial explosion in the number of possible spatiotemporal steps, the optimization routine is intractable and requires high-performance computing (HPC) strategies to solve realistic systems.
The presentation will summarize industrial benchmark results where A100 GPUs accelerate training and inference aspects of reinforcement learning, yielding significant performance improvements and ideal NPV solutions. Finally, Vidyasagar will outline the future of deep RL in the context of HPC in energy.