skip to main content

Title: Explaining the physics of transfer learning in data-driven turbulence modeling

Transfer learning (TL), which enables neural networks (NNs) to generalize out-of-distribution via targeted re-training, is becoming a powerful tool in scientific machine learning (ML) applications such as weather/climate prediction and turbulence modeling. Effective TL requires knowing (1) how to re-train NNs? and (2) what physics are learned during TL? Here, we present novel analyses and a framework addressing (1)–(2) for a broad range of multi-scale, nonlinear, dynamical systems. Our approach combines spectral (e.g. Fourier) analyses of such systems with spectral analyses of convolutional NNs, revealing physical connections between the systems and what the NN learns (a combination of low-, high-, band-pass filters and Gabor filters). Integrating these analyses, we introduce a general framework that identifies the best re-training procedure for a given problem based on physics and NN theory. As test case, we explain the physics of TL in subgrid-scale modeling of several setups of 2D turbulence. Furthermore, these analyses show that in these cases, the shallowest convolution layers are the best to re-train, which is consistent with our physics-guided framework but is against the common wisdom guiding TL in the ML literature. Our work provides a new avenue for optimal and explainable TL, and a step toward fully explainable NNs, for wide-ranging applications in science and engineering, such as climate change modeling.

more » « less
Award ID(s):
Author(s) / Creator(s):
; ; ;
Yortsos, Yannis
Publisher / Repository:
Oxford University
Date Published:
Journal Name:
PNAS Nexus
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Metal–organic frameworks (MOFs) are promising materials with various applications, and machine learning (ML) techniques can enable their design and understanding of structure–property relationships. In this paper, we use machine learning (ML) to cluster the MOFs using two different approaches. For the first set of clusters, we decompose the data using the textural properties and cluster the resulting components. We separately cluster the MOF space with respect to their topology. The feature data from each of the clusters were then fed into separate neural networks (NNs) for direct learning on an adsorption task (methane or hydrogen). The resulting NNs were then used in transfer learning (TL) where only the last NN layer was retrained. The results show significant differences in TL performance based on which cluster is chosen for direct learning. We find TL performance depends on the Euclidean distance in the decomposed feature space between the clusters involved in the direct and TL. Similar results were found when TL was performed simultaneously across both types of clusters and adsorption tasks. We note that methane adsorption was a better source task than hydrogen adsorption. Overall, the approach was able to identify MOFs with the most transferable information, leading to valuable insights and a more comprehensive understanding of the MOF landscape. This highlights the method's potential to generate a deeper understanding of complex systems and provides an opportunity for its application in alternative datasets. 
    more » « less
  2. null (Ed.)
    Artificial neural networks (NNs) in deep learning systems are critical drivers of emerging technologies such as computer vision, text classification, and natural language processing. Fundamental to their success is the development of accurate and efficient NN models. In this article, we report our work on Deep-n-Cheap—an open-source automated machine learning (AutoML) search framework for deep learning models. The search includes both architecture and training hyperparameters and supports convolutional neural networks and multi-layer perceptrons, applicable to multiple domains. Our framework is targeted for deployment on both benchmark and custom datasets, and as a result, offers a greater degree of search space customizability as compared to a more limited search over only pre-existing models from literature. We also introduce the technique of ‘search transfer’, which demonstrates the generalization capabilities of the models found by our framework to multiple datasets. Deep-n-Cheap includes a user-customizable complexity penalty which trades off performance with training time or number of parameters. Specifically, our framework can find models with performance comparable to state-of-the- art while taking 1–2 orders of magnitude less time to train than models from other AutoML and model search frameworks. Additionally, we investigate and develop insight into the search process that should aid future development of deep learning models. 
    more » « less
  3. Abstract

    In the last decade, much work in atmospheric science has focused on spatial verification (SV) methods for gridded prediction, which overcome serious disadvantages of pixelwise verification. However, neural networks (NN) in atmospheric science are almost always trained to optimize pixelwise loss functions, even when ultimately assessed with SV methods. This establishes a disconnect between model verification during versus after training. To address this issue, we develop spatially enhanced loss functions (SELF) and demonstrate their use for a real-world problem: predicting the occurrence of thunderstorms (henceforth, “convection”) with NNs. In each SELF we use either a neighborhood filter, which highlights convection at scales larger than a threshold, or a spectral filter (employing Fourier or wavelet decomposition), which is more flexible and highlights convection at scales between two thresholds. We use these filters to spatially enhance common verification scores, such as the Brier score. We train each NN with a different SELF and compare their performance at many scales of convection, from discrete storm cells to tropical cyclones. Among our many findings are that (i) for a low or high risk threshold, the ideal SELF focuses on small or large scales, respectively; (ii) models trained with a pixelwise loss function perform surprisingly well; and (iii) nevertheless, models trained with a spectral filter produce much better-calibrated probabilities than a pixelwise model. We provide a general guide to using SELFs, including technical challenges and the final Python code, as well as demonstrating their use for the convection problem. To our knowledge this is the most in-depth guide to SELFs in the geosciences.

    Significance Statement

    Gridded predictions, in which a quantity is predicted at every pixel in space, should be verified with spatially aware methods rather than pixel by pixel. Neural networks (NN), which are often used for gridded prediction, are trained to minimize an error value called the loss function. NN loss functions in atmospheric science are almost always pixelwise, which causes the predictions to miss rare events and contain unrealistic spatial patterns. We use spatial filters to enhance NN loss functions, and we test our novel spatially enhanced loss functions (SELF) on thunderstorm prediction. We find that different SELFs work better for different scales (i.e., different-sized thunderstorm complexes) and that spectral filters, one of the two filter types, produce unexpectedly well calibrated thunderstorm probabilities.

    more » « less
  4. Abstract

    Methods of explainable artificial intelligence (XAI) are used in geoscientific applications to gain insights into the decision-making strategy of neural networks (NNs), highlighting which features in the input contribute the most to a NN prediction. Here, we discuss our “lesson learned” that the task of attributing a prediction to the input does not have a single solution. Instead, the attribution results depend greatly on the considered baseline that the XAI method utilizes—a fact that has been overlooked in the geoscientific literature. The baseline is a reference point to which the prediction is compared so that the prediction can be understood. This baseline can be chosen by the user or is set by construction in the method’s algorithm—often without the user being aware of that choice. We highlight that different baselines can lead to different insights for different science questions and, thus, should be chosen accordingly. To illustrate the impact of the baseline, we use a large ensemble of historical and future climate simulations forced with the shared socioeconomic pathway 3-7.0 (SSP3-7.0) scenario and train a fully connected NN to predict the ensemble- and global-mean temperature (i.e., the forced global warming signal) given an annual temperature map from an individual ensemble member. We then use various XAI methods and different baselines to attribute the network predictions to the input. We show that attributions differ substantially when considering different baselines, because they correspond to answering different science questions. We conclude by discussing important implications and considerations about the use of baselines in XAI research.

    Significance Statement

    In recent years, methods of explainable artificial intelligence (XAI) have found great application in geoscientific applications, because they can be used to attribute the predictions of neural networks (NNs) to the input and interpret them physically. Here, we highlight that the attributions—and the physical interpretation—depend greatly on the choice of the baseline—a fact that has been overlooked in the geoscientific literature. We illustrate this dependence for a specific climate task, in which a NN is trained to predict the ensemble- and global-mean temperature (i.e., the forced global warming signal) given an annual temperature map from an individual ensemble member. We show that attributions differ substantially when considering different baselines, because they correspond to answering different science questions.

    more » « less
  5. The water content in the soil regulates exchanges between soil and atmosphere, impacts plant livelihood, and determines the antecedent condition for several natural hazards. Accurate soil moisture estimates are key to applications such as natural hazard prediction, agriculture, and water management. We explore how to best predict soil moisture at a high resolution in the context of a changing climate. Physics-based hydrological models are promising as they provide distributed soil moisture estimates and allow prediction outside the range of prior observations. This is particularly important considering that the climate is changing, and the available historical records are often too short to capture extreme events. Unfortunately, these models are extremely computationally expensive, which makes their use challenging, especially when dealing with strong uncertainties. These characteristics make them complementary to machine learning approaches, which rely on training data quality/quantity but are typically computationally efficient. We first demonstrate the ability of Convolutional Neural Networks (CNNs) to reproduce soil moisture fields simulated by the hydrological model ParFlow-CLM. Then, we show how these two approaches can be successfully combined to predict future droughts not seen in the historical timeseries. We do this by generating additional ParFlow-CLM simulations with altered forcing mimicking future drought scenarios. Comparing the performance of CNN models trained on historical forcing and CNN models trained also on simulations with altered forcing reveals the potential of combining these two approaches. The CNN can not only reproduce the moisture response to a given forcing but also learn and predict the impact of altered forcing. Given the uncertainties in projected climate change, we can create a limited number of representative ParFlow-CLM simulations (ca. 25 min/water year on 9 CPUs for our case study), train our CNNs, and use them to efficiently (seconds/water-year on 1 CPU) predict additional water years/scenarios and improve our understanding of future drought potential. This framework allows users to explore scenarios beyond past observation and tailor the training data to their application of interest (e.g., wet conditions for flooding, dry conditions for drought, etc…). With the trained ML model they can rely on high resolution soil moisture estimates and explore the impact of uncertainties.

    more » « less