Back in 2016, in his editorial ‘Watson and the weather’, Oil IT Journal editor Neil McNaughton speculated on the possibility that weather forecasting might prove to be a litmus test for artificial intelligence. The huge data sets acquired over decades ought to provide a great comparison between full physics based forward modeling and data driven prediction. It turns out that such questions are at the cutting edge of climate research, as a recent release from the Department of Energy’s Lawrence Berkeley National Laboratory showed. A Berkeley Lab team is developing ClimateNet to ‘bring the power’ of deep learning methods to identify important weather and climate patterns via expert-labeled, community-sourced open datasets and architectures. The release has it that the resulting deep learning models will identify complex weather and climate patterns on a global scale.
We spotted an opportunity to take a raincheck on the state of play in the match between deep learning and physics-based modeling and asked Berkeley researcher Prabhat why a data driven model might do better than a traditional, full physics-based forward model, and whether this would be true for other industries that use computer modeling, like oil and gas. Prabhat first pointed out that ClimateNet is purely aimed at post-processing climate model output (i.e. finding patterns in simulation data) and is not proposing to replace a climate modeling with a data-driven ML/DL approach. He added, ‘When comparing data-driven ML/DL models to physics-driven models, we don’t have sufficient evidence to prove that ML will be universally better. If anything, ML methods will suffer from access to limited training data and may not generalize to regimes that have not been seen before’. We then invited Prabhat to read our 2016 editorial and comment which he kindly did.
This is probably a much longer conversation, but some quick observations:
Within the DOE, we’ve identified three broad areas for Deep Learning/Machine Learning: 1) DL for Data Analytics (think analysis of datasets from telescopes, microscopes, genome sequencers, etc. Much of ‘Big Data Analytics’ technologies are targeting this space. 2) DL for Simulations this is about augmenting, enhancing and *maybe* replacing conventional PDE*-based solvers which your article refers to this as ‘forward’ models 3) DL for control this is about controlling experimental facilities (telescopes, light source beam lines) or computational facilities.
Your article was speculating about whether data driven methods (DL/ML/Watson/…) could replace conventional models. Clearly, this is a hot topic for research, and there aren’t conclusive trends at this point. We understand very little about the theory and generalization properties of DL, we don’t have ‘proofs’, training the DL system on a certain climate regime, and asking it to make predictions in another climate regime could be problematic.
It is becoming clear that we shouldn’t throw away 40+ years of applied math research. Incorporating some notion of physics laws/principles into black-box DL models will likely be key. For more on this checkout the paper below**. We also have a major paper under review (for the Gordon Bell Prize) that employs Deep Learning to learn solutions of stochastic differential equations.
* Partial differential equation.
** Deep Hidden Physics Models: Deep Learning of Nonlinear Partial Differential Equations, Maziar Raissi 2018 - https://arxiv.org/pdf/1801.06637.
In a possibly similar vein, we came across a recent NIST investigation into Explainability in Artificial Intelligence and Machine Learning in a computer security context.
© Oil IT Journal - all rights reserved.