Search NASA⌕ Search

SEARCH · Search NASA

Results for “Biological neural networks”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 307 records · Page 17
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

Ultrafast radiographic imaging and tracking: An overview of instruments, methods, data, and applications

Ultrafast radiographic imaging and tracking (U-RadIT) use state-of-the-art ionizing particle and light sources to experimentally study sub-nanosecond transients or dynamic processes in physics, chemistry, biology, geology, materials science and other fields. These processes are fundamental to modern technologies and applications, such as nuclear fusion energy, advanced manufacturing, communication, and green transportation, which often involve one mole or more atoms and elementary particles, and thus are challenging to compute by using the first principles of quantum physics or other forward models. One of the central problems in U-RadIT is to optimize information yield through, e.g. high-luminosity X-ray and particle sources, efficient imaging and tracking detectors, novel methods to collect data, and large-bandwidth online and offline data processing, regulated by the underlying physics, statistics, and computing power. We review and highlight recent progress in: (a.) Detectors such as high-speed complementary metal-oxide semiconductor (CMOS) cameras, hybrid pixelated array detectors integrated with Timepix4 and other application-specific integrated circuits (ASICs), and digital photon detectors; (b.) U-RadIT modalities such as dynamic phase contrast imaging, dynamic diffractive imaging, and four-dimensional (4D) particle tracking; (c.) U-RadIT data and algorithms such as neural networks and machine learning, and (d.) Applications in ultrafast dynamic material science using XFELs, synchrotrons and laser-driven sources. Hardware-centric approaches to U-RadIT optimization are constrained by detector material properties, low signal-to-noise ratio, high cost and long development cycles of critical hardware components such as ASICs. Interpretation of experimental data, including comparisons with forward models, is frequently hindered by sparse measurements, model and measurement uncertainties, and noise. Alternatively, U-RadIT make increasing use of data science and machine learning algorithms, including experimental implementations of compressed sensing. Machine learning and artificial intelligence approaches, refined by physics and materials information, may also contribute significantly to data interpretation, uncertainty quantification and U-RadIT optimization.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2022

Scaling and Benchmarking an Evolutionary Algorithm for Constructing Biophysical Neuronal Models

Single neuron models are fundamental for computational modeling of the brain's neuronal networks, and understanding how ion channel dynamics mediate neural function. A challenge in defining such models is determining biophysically realistic channel distributions. Here, we present an efficient, highly parallel evolutionary algorithm for developing such models, named NeuroGPU-EA. NeuroGPU-EA uses CPUs and GPUs concurrently to simulate and evaluate neuron membrane potentials with respect to multiple stimuli. We demonstrate a logarithmic cost for scaling the stimuli used in the fitting procedure. NeuroGPU-EA outperforms the typically used CPU based evolutionary algorithm by a factor of 10 on a series of scaling benchmarks. We report observed performance bottlenecks and propose mitigation strategies. Finally, we also discuss the potential of this method for efficient simulation and evaluation of electrophysiological waveforms.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2023

Identification of mobile genetic elements with geNomad

Identifying and characterizing mobile genetic elements in sequencing data is essential for understanding their diversity, ecology, biotechnological applications and impact on public health. Here we introduce geNomad, a classification and annotation framework that combines information from gene content and a deep neural network to identify sequences of plasmids and viruses. geNomad uses a dataset of more than 200,000 marker protein profiles to provide functional gene annotation and taxonomic assignment of viral genomes. Using a conditional random field model, geNomad also detects proviruses integrated into host genomes with high precision. In benchmarks, geNomad achieved high classification performance for diverse plasmids and viruses (Matthews correlation coefficient of 77.8% and 95.3%, respectively), substantially outperforming other tools. Leveraging geNomad’s speed and scalability, we processed over 2.7 trillion base pairs of sequencing data, leading to the discovery of millions of viruses and plasmids that are available through the IMG/VR and IMG/PR databases. geNomad is available at https://portal.nersc.gov/genomad.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2021

Transfer Learning using Denoising Auto-Encoders for Cellular-Level Annotation of Tumor in Pathology Slides

Adversarial examples can produce altered classifications using only seemingly innocuous, imperceptible perturbations to the original image. The imperceptibility of adversarial perturbations suggests that the corresponding classifiers use decision criteria different than those of a human. In a medical setting, inexplicable decision criteria confound a pathologist’s willingness to trust machine-generated annotations. Here, we analyze denoising tumor detection models to see if they are robust to imperceptible adversarial perturbations. Moreover, to be more fully trusted by pathologists, we require tumor detectors that generate interpretable annotations which segment pathology slides into tumorous and normal regions at the cellular level. We therefore compare transfer learning based on two different autoencoder architectures, one derived from a deep denoising bottleneck autoencoder and one from an over-complete sparse autoencoder. Both autoencoders were first trained in an unsupervised manner on a set of pathology slides drawn from the Camelyon16 dataset. The latent representations produced by each autoencoder were then passed to separate neural networks that were trained in a supervised manner on binary tumor-normal masks generated by pathologists at cellular resolution. Both tumor detectors supported better than 90% AUC PR as measured by the area under the precision/recall curve on a held-out pathology slide. To assess the underlying decision criteria used by both tumor detectors, we constructed imperceptible adversarial examples which reduced the AUC PR of both models to less than 70%. Random noise of the same amplitude had almost no effect on the AUC PR of either model. Additionally, each tumor detector was resistant to adversarial “transfer” attacks targeting the other. The adversarial perturbations showed strong characteristic differences: the deep denoising models perturbations were a very diffuse, seemingly unrecognizable pattern while the sparse coding models perturbations showed traces of tissue cells.

47 OTHER INSTRUMENTATION↗
DOE OSTI2021

AutoAtlas: Neural Network for 3D Unsupervised Partitioning and Representation Learning

Here we present a novel neural network architecture called AutoAtlas for fully unsupervised partitioning and representation learning of 3D brain Magnetic Resonance Imaging (MRI) volumes. AutoAtlas consists of two neural network components: one neural network to perform multi-label partitioning based on local texture in the volume, and a second neural network to compress the information contained within each partition. We train both of these components simultaneously by optimizing a loss function that is designed to promote accurate reconstruction of each partition, while encouraging spatially smooth and contiguous partitioning, and discouraging relatively small partitions. We show that the partitions adapt to the subject specific structural variations of brain tissue while consistently appearing at similar spatial locations across subjects. AutoAtlas also produces very low dimensional features that represent local texture of each partition. We demonstrate prediction of metadata associated with each subject using the derived feature representations and compare the results to prediction using features derived from FreeSurfer anatomical parcellation. Since our features are intrinsically linked to distinct partitions, we can then map values of interest, such as partition-specific feature importance scores onto the brain for visualization.

42 ENGINEERING↗
DOE OSTI2022

High–throughput measurement of plant fitness traits with an object detection method using Faster R–CNN

Revealing the contributions of genes to plant phenotype is frequently challenging because loss-of-function effects may be subtle or masked by varying degrees of genetic redundancy. Such effects can potentially be detected by measuring plant fitness, which reflects the cumulative effects of genetic changes over the lifetime of a plant. However, fitness is challenging to measure accurately, particularly in species with high fecundity and relatively small propagule sizes such as Arabidopsis thaliana. An image segmentation-based method using the software ImageJ and an object detection-based method using the Faster Region-based Convolutional Neural Network (R-CNN) algorithm were used for measuring two Arabidopsis fitness traits: seed and fruit counts. The segmentation-based method was error-prone (correlation between true and predicted seed counts, r 2 = 0.849) because seeds touching each other were undercounted. By contrast, the object detection-based algorithm yielded near perfect seed counts (r 2 = 0.9996) and highly accurate fruit counts (r 2 = 0.980). Comparing seed counts for wild-type and 12 mutant lines revealed fitness effects for three genes; fruit counts revealed the same effects for two genes. Our study provides analysis pipelines and models to facilitate the investigation of Arabidopsis fitness traits and demonstrates the importance of examining fitness traits when studying gene functions.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2020

Enhancing Streamflow Forecast and Extracting Insights Using Long-Short Term Memory Networks With Data Integration at Continental Scales

Recent observations with varied schedules and types (moving average, snapshot, or regularly spaced) can help to improve streamflow forecasts, but it is challenging to integrate them effectively. Based on a long short-term memory (LSTM) streamflow model, we tested multiple versions of a flexible procedure we call data integration (DI) to leverage recent discharge measurements to improve forecasts. DI accepts lagged inputs either directly or through a convolutional neural network unit. DI ubiquitously elevated streamflow forecast performance to unseen levels, reaching a record continental-scale median Nash-Sutcliffe Efficiency coefficient value of 0.86. Integrating moving-average discharge, discharge from the last few days, or even average discharge from the previous calendar month could all improve daily forecasts. Directly using lagged observations as inputs was comparable in performance to using the convolutional neural network unit. Importantly, we obtained valuable insights regarding hydrologic processes impacting LSTM and DI performance. Before applying DI, the base LSTM model worked well in mountainous or snow-dominated regions, but less well in regions with low discharge volumes (due to either low precipitation or high precipitation-energy synchronicity) and large interannual storage variability. DI was most beneficial in regions with high flow autocorrelation: it greatly reduced baseflow bias in groundwater-dominated western basins and also improved peak prediction for basins with dynamical surface water storage, such as the Prairie Potholes or Great Lakes regions. However, even DI cannot elevate performance in high-aridity basins with 1-day flash peaks. Despite this limitation, there is much promise for a deep-learning-based forecast paradigm due to its performance, automation, efficiency, and flexibility.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2023

A differentiable, physics-informed ecosystem modeling and learning framework for large-scale inverse problems: demonstration with photosynthesis simulations

Photosynthesis plays an important role in carbon, nitrogen, and water cycles. Ecosystem models for photosynthesis are characterized by many parameters that are obtained from limited in situ measurements and applied to the same plant types. Previous site-by-site calibration approaches could not leverage big data and faced issues like overfitting or parameter non-uniqueness. Here we developed an end-to-end programmatically differentiable (meaning gradients of outputs to variables used in the model can be obtained efficiently and accurately) version of the photosynthesis process representation within the Functionally Assembled Terrestrial Ecosystem Simulator (FATES) model. As a genre of physics-informed machine learning (ML), differentiable models couple physics-based formulations to neural networks (NNs) that learn parameterizations (and potentially processes) from observations, here photosynthesis rates. We first demonstrated that the framework was able to correctly recover multiple assumed parameter values concurrently using synthetic training data. Then, using a real-world dataset consisting of many different plant functional types (PFTs), we learned parameters that performed substantially better and greatly reduced biases compared to literature values. Further, the framework allowed us to gain insights at a large scale. Our results showed that the carboxylation rate at 25 °C (V c,max25 ) was more impactful than a factor representing water limitation, although tuning both was helpful in addressing biases with the default values. This framework could potentially enable substantial improvement in our capability to learn parameters and reduce biases for ecosystem modeling at large scales.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

DISTEMA: distance map-based estimation of single protein model accuracy with attentive 2D convolutional neural network

Abstract Background Estimation of the accuracy (quality) of protein structural models is important for both prediction and use of protein structural models. Deep learning methods have been used to integrate protein structure features to predict the quality of protein models. Inter-residue distances are key information for predicting protein’s tertiary structures and therefore have good potentials to predict the quality of protein structural models. However, few methods have been developed to fully take advantage of predicted inter-residue distance maps to estimate the accuracy of a single protein structural model. Result We developed an attentive 2D convolutional neural network (CNN) with channel-wise attention to take only a raw difference map between the inter-residue distance map calculated from a single protein model and the distance map predicted from the protein sequence as input to predict the quality of the model. The network comprises multiple convolutional layers, batch normalization layers, dense layers, and Squeeze-and-Excitation blocks with attention to automatically extract features relevant to protein model quality from the raw input without using any expert-curated features. We evaluated DISTEMA’s capability of selecting the best models for CASP13 targets in terms of ranking loss of GDT-TS score. The ranking loss of DISTEMA is 0.079, lower than several state-of-the-art single-model quality assessment methods. Conclusion This work demonstrates that using raw inter-residue distance information with deep learning can predict the quality of protein structural models reasonably well. DISTEMA is freely at https://github.com/jianlin-cheng/DISTEMA

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2023

Multimodal representation learning for predicting molecule–disease relations

Motivation: Predicting molecule–disease indications and side effects is important for drug development and pharmacovigilance. Comprehensively mining molecule–molecule, molecule–disease and disease–disease semantic dependencies can potentially improve prediction performance. Methods: We introduce a Multi-Modal REpresentation Mapping Approach to Predicting molecular-disease relations (M2REMAP) by incorporating clinical semantics learned from electronic health records (EHR) of 12.6 million patients. Specifically, M2REMAP first learns a multimodal molecule representation that synthesizes chemical property and clinical semantic information by mapping molecule chemicals via a deep neural network onto the clinical semantic embedding space shared by drugs, diseases and other common clinical concepts. To infer molecule–disease relations, M2REMAP combines multimodal molecule representation and disease semantic embedding to jointly infer indications and side effects. Results: We extensively evaluate M2REMAP on molecule indications, side effects and interactions. Results show that incorporating EHR embeddings improves performance significantly, for example, attaining an improvement over the baseline models by 23.6% in PRC-AUC on indications and 23.9% on side effects. Further, M2REMAP overcomes the limitation of existing methods and effectively predicts drugs for novel diseases and emerging pathogens. Availability and implementation: The code is available at https://github.com/celehs/M2REMAP, and prediction results are provided at https://shiny.parse-health.org/drugs-diseases-dev/.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2023

DL-TODA: A Deep Learning Tool for Omics Data Analysis

Metagenomics is a technique for genome-wide profiling of microbiomes; this technique generates billions of DNA sequences called reads. Given the multiplication of metagenomic projects, computational tools are necessary to enable the efficient and accurate classification of metagenomic reads without needing to construct a reference database. The program DL-TODA presented here aims to classify metagenomic reads using a deep learning model trained on over 3000 bacterial species. A convolutional neural network architecture originally designed for computer vision was applied for the modeling of species-specific features. Using synthetic testing data simulated with 2454 genomes from 639 species, DL-TODA was shown to classify nearly 75% of the reads with high confidence. The classification accuracy of DL-TODA was over 0.98 at taxonomic ranks above the genus level, making it comparable with Kraken2 and Centrifuge, two state-of-the-art taxonomic classification tools. DL-TODA also achieved an accuracy of 0.97 at the species level, which is higher than 0.93 by Kraken2 and 0.85 by Centrifuge on the same test set. Application of DL-TODA to the human oral and cropland soil metagenomes further demonstrated its use in analyzing microbiomes from diverse environments. Compared to Centrifuge and Kraken2, DL-TODA predicted distinct relative abundance rankings and is less biased toward a single taxon.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2021

ECNet is an evolutionary context-integrated deep learning framework for protein engineering

Abstract Machine learning has been increasingly used for protein engineering. However, because the general sequence contexts they capture are not specific to the protein being engineered, the accuracy of existing machine learning algorithms is rather limited. Here, we report ECNet (evolutionary context-integrated neural network), a deep-learning algorithm that exploits evolutionary contexts to predict functional fitness for protein engineering. This algorithm integrates local evolutionary context from homologous sequences that explicitly model residue-residue epistasis for the protein of interest with the global evolutionary context that encodes rich semantic and structural features from the enormous protein sequence universe. As such, it enables accurate mapping from sequence to function and provides generalization from low-order mutants to higher-order mutants. We show that ECNet predicts the sequence-function relationship more accurately as compared to existing machine learning algorithms by using ~50 deep mutational scanning and random mutagenesis datasets. Moreover, we used ECNet to guide the engineering of TEM-1 β-lactamase and identified variants with improved ampicillin resistance with high success rates.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2022

An image classification deep-learning algorithm for shrapnel detection from ultrasound images

Ultrasound imaging is essential for non-invasively diagnosing injuries where advanced diagnostics may not be possible. However, image interpretation remains a challenge as proper expertise may not be available. In response, artificial intelligence algorithms are being investigated to automate image analysis and diagnosis. Here, we highlight an image classification convolutional neural network for detecting shrapnel in ultrasound images. As an initial application, different shrapnel types and sizes were embedded first in a tissue mimicking phantom and then in swine thigh tissue. The algorithm architecture was optimized stepwise by minimizing validation loss and maximizing F1 score. The final algorithm design trained on tissue phantom image sets had an F1 score of 0.95 and an area under the ROC curve of 0.95. It maintained higher than a 90% accuracy for each of 8 shrapnel types. When trained only on swine image sets, the optimized algorithm format had even higher metrics: F1 and area under the ROC curve of 0.99. Overall, the algorithm developed resulted in strong classification accuracy for both the tissue phantom and animal tissue. This framework can be applied to other trauma relevant imaging applications such as internal bleeding to further simplify trauma medicine when resources and image interpretation are scarce.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2021

Global transcriptome profile of the developmental principles of in vitro iPSC-to-motor neuron differentiation

Abstract Background Human induced pluripotent stem cells (iPSC) have opened new avenues for regenerative medicine. Consequently, iPSC-derived motor neurons have emerged as potentially viable therapies for spinal cord injuries and neurodegenerative disorders including Amyotrophic Lateral Sclerosis. However, direct clinical application of iPSC bears in itself the risk of tumorigenesis and other unforeseeable genetic or epigenetic abnormalities. Results Employing RNA-seq technology, we identified and characterized gene regulatory networks triggered by in vitro chemical reprogramming of iPSC into cells with the molecular features of motor neurons (MNs) whose function in vivo is to innervate effector organs. We present meta-transcriptome signatures of 5 cell types: iPSCs, neural stem cells, motor neuron progenitors, early motor neurons, and mature motor neurons. In strict response to the chemical stimuli, along the MN differentiation axis we observed temporal downregulation of tumor growth factor-β signaling pathway and consistent activation of sonic hedgehog, Wnt/β-catenin, and Notch signaling. Together with gene networks defining neuronal differentiation (neurogenin 2, microtubule-associated protein 2, Pax6, and neuropilin-1), we observed steady accumulation of motor neuron-specific regulatory genes, including Islet-1 and homeobox protein HB9. Interestingly, transcriptome profiling of the differentiation process showed that Ca 2+ signaling through cAMP and LPC was downregulated during the conversion of the iPSC to neural stem cells and key regulatory gene activity of the pathway remained inhibited until later stages of motor neuron formation. Pathways shaping the neuronal development and function were well-represented in the early motor neuron cells including, neuroactive ligand-receptor interactions, axon guidance, and the cholinergic synapse formation. A notable hallmark of our in vitro motor neuron maturation in monoculture was the activation of genes encoding G-coupled muscarinic acetylcholine receptors and downregulation of the ionotropic nicotinic acetylcholine receptors expression. We observed the formation of functional neuronal networks as spontaneous oscillations in the extracellular action potentials recorded on multi-electrode array chip after 20 days of differentiation. Conclusions Detailed transcriptome profile of each developmental step from iPSC to motor neuron driven by chemical induction provides the guidelines to novel therapeutic approaches in the re-construction efforts of muscle innervation.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2020

Deeply uncertain: comparing methods of uncertainty quantification in deep learning algorithms

We present a comparison of methods for uncertainty quantification (UQ) in deep learning algorithms in the context of a simple physical system. Three of the most common uncertainty quantification methods - Bayesian Neural Networks (BNN), Concrete Dropout (CD), and Deep Ensembles (DE) - are compared to the standard analytic error propagation. We discuss this comparison in terms endemic to both machine learning ("epistemic" and "aleatoric") and the physical sciences ("statistical" and "systematic"). The comparisons are presented in terms of simulated experimental measurements of a single pendulum - a prototypical physical system for studying measurement and analysis techniques. Our results highlight some pitfalls that may occur when using these UQ methods. For example, when the variation of noise in the training set is small, all methods predicted the same relative uncertainty independently of the inputs. This issue is particularly hard to avoid in BNN. On the other hand, when the test set contains samples far from the training distribution, we found that no methods sufficiently increased the uncertainties associated to their predictions. This problem was particularly clear for CD. In light of these results, we make some recommendations for usage and interpretation of UQ methods.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2020

Overcoming small minirhizotron datasets using transfer learning

Minirhizotron technology is widely used to study root growth and development. Yet, standard approaches for tracing roots in minirhiztron imagery is extremely tedious and time consuming. Machine learning approaches can help to automate this task. However, lack of enough annotated training data is a major limitation for the application of machine learning methods. Transfer learning is a useful technique to help with training when available datasets are limited. In this paper, we investigated the effect of pre-trained features from the massives-cale, irrelevant ImageNet dataset and a relatively moderate-scale, but relevant peanut root dataset on switchgrass root imagery segmentation applications. We compiled two minirhizotron image datasets to accomplish this study: one with 17,550 peanut root images and another with 28 switchgrass root images. Both datasets were paired with manually labeled ground truth masks. Deep neural networks based on the U-net architecture were used with different pre-trained features as initialization for automated, precise pixel-wise root segmentation in minirhizotron imagery. We observed that features pre-trained on a closely related but relatively moderate size dataset like our peanut dataset were more effective than features pre-trained on the large but unrelated ImageNet dataset. Here, we achieved high quality segmentation on peanut root dataset with 99.04% accuracy at the pixel-level and overcame errors in human-labeled ground truth masks. By applying transfer learning technique on limited switchgrass dataset with features pre-trained on peanut dataset, we obtained 99% segmentation accuracy in switchgrass imagery using only 21 images for training (fine tuning). Furthermore, the peanut pre-trained features can help the model converge faster and have much more stable performance.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2024

Computer vision models enable mixed linear modeling to predict arbuscular mycorrhizal fungal colonization using fungal morphology

Abstract The presence of Arbuscular Mycorrhizal Fungi (AMF) in vascular land plant roots is one of the most ancient of symbioses supporting nitrogen and phosphorus exchange for photosynthetically derived carbon. Here we provide a multi-scale modeling approach to predict AMF colonization of a worldwide crop from a Recombinant Inbred Line (RIL) population derived from Sorghum bicolor and S. propinquum . The high-throughput phenotyping methods of fungal structures here rely on a Mask Region-based Convolutional Neural Network (Mask R-CNN) in computer vision for pixel-wise fungal structure segmentations and mixed linear models to explore the relations of AMF colonization, root niche, and fungal structure allocation. Models proposed capture over 95% of the variation in AMF colonization as a function of root niche and relative abundance of fungal structures in each plant. Arbuscule allocation is a significant predictor of AMF colonization among sibling plants. Arbuscules and extraradical hyphae implicated in nutrient exchange predict highest AMF colonization in the top root section. Our work demonstrates that deep learning can be used by the community for the high-throughput phenotyping of AMF in plant roots. Mixed linear modeling provides a framework for testing hypotheses about AMF colonization phenotypes as a function of root niche and fungal structure allocations.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2022

Differentiable, Learnable, Regionalized Process-Based Models With Multiphysical Outputs can Approach State-Of-The-Art Hydrologic Prediction Accuracy

Predictions of hydrologic variables across the entire water cycle have significant value for water resources management as well as downstream applications such as ecosystem and water quality modeling. Recently, purely data-driven deep learning models like long short-term memory (LSTM) showed seemingly insurmountable performance in modeling rainfall runoff and other geoscientific variables, yet they cannot predict untrained physical variables and remain challenging to interpret. Here, we show that differentiable, learnable, process-based models (called δ models here) can approach the performance level of LSTM for the intensively observed variable (streamflow) with regionalized parameterization. We use a simple hydrologic model HBV as the backbone and use embedded neural networks, which can only be trained in a differentiable programming framework, to parameterize, enhance, or replace the process-based model's modules. Without using an ensemble or post-processor, δ models can obtain a median Nash-Sutcliffe efficiency of 0.732 for 671 basins across the USA for the Daymet forcing data set, compared to 0.748 from a state-of-the-art LSTM model with the same setup. For another forcing data set, the difference is even smaller: 0.715 versus 0.722. Meanwhile, the resulting learnable process-based models can output a full set of untrained variables, for example, soil and groundwater storage, snowpack, evapotranspiration, and baseflow, and can later be constrained by their observations. Both simulated evapotranspiration and fraction of discharge from baseflow agreed decently with alternative estimates. The general framework can work with models with various process complexity and opens up the path for learning physics from big data.

54 ENVIRONMENTAL SCIENCES↗