Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “learned”

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 541 records · Page 30

Transfer learning-based soybean LAI estimations by integrating PROSAIL, UAV, and PlanetScope imagery

Accurate Leaf Area Index (LAI) estimations at the soybean plot scale is achievable using high-resolution Unmanned Aerial Vehicle (UAV) imagery and field measurement samples. However, the limited coverage of UAV flights restricts large-scale remote sensing monitoring in expansive soybean fields. This study leverages the broad coverage and 3-m resolution of PlanetScope satellite imagery to extend LAI prediction from UAV to satellite scales through transfer learning, using UAV-scale LAI estimates as a benchmark to validate cross-scale consistency. To address this challenge, this study proposed the LAI-TransNet, a two-stage transfer learning framework designed for precise and scalable soybean LAI prediction across large areas, demonstrating its effectiveness in cross-scale monitoring. In Stage 1, a UAV-scale benchmark is established using PROSAIL-simulated UAV reflectance data (UAV-Sim) and field-measured soybean LAI. Traditional machine learning, deep learning, and transfer learning models are trained on a hybrid UAV-Sim and field-measured dataset (UAV-Sim_Measured), with the transfer learning model CNN-TL, fine-tuned using pre-trained weights derived from UAV-Sim, achieving the highest accuracy (R 2 = 0.81, RMSE = 0.64 m 2 /m 2 , rRMSE = 11.5 %). In Stage 2, LAI-TransNet is developed by fine-tuning the CNN-TL model on PlanetScope simulated data (PS-Sim), preprocessed via cross-domain mapping to align UAV and satellite spectral features. Real PlanetScope imagery is corrected for reflectance consistency with reference to UAV imagery spectral profiles. LAI-TransNet outperforms other deep learning models trained directly on PS-Sim (R 2 = 0.69 vs. 0.60–0.63), ensuring robust cross-scale consistency. In conclusion, by bridging UAV and satellite scales, LAI-TransNet enables large-scale soybean LAI monitoring, enhancing precision agriculture management through improved monitoring with the PlanetScope imagery.

Leaf area index (LAI)↗

Learning through EC directive based SEA in spatial planning? Evidence from the Brunswick Region in Germany

This paper presents results of an international comparative research project, funded by the UK Economic and Social Research Council (ESRC) and the Academy for Sustainable Communities (ASC) on the 'learning potential of appraisal (strategic environmental assessment - SEA) in spatial planning'. In this context, aspects of 'single-loop' and 'double-loop' learning, as well as of individual, organisational and social learning are discussed for emerging post-EC Directive German practice in the planning region (Zweckverband) of Brunswick (Braunschweig), focusing on four spatial plan SEAs from various administrative levels in the region. It is found that whilst SEA is able to lead to plan SEA specific knowledge acquisition, comprehension, application and analysis ('single-loop learning'), it is currently resulting only occasionally in wider synthesis and evaluation ('double-loop learning'). Furthermore, whilst there is evidence that individual and occasionally organisational learning may be enhanced through SEA, most notably in small municipalities, social learning appears to be happening only sporadically.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗

Public participation in and learning through SEA in Kenya

Meaningful public engagement is a challenging, but promising, feature of strategic environmental assessment (SEA) due to its potential for integrating sustainability principles into policies, plans and programs in developing countries such as Kenya. This research examined two selected SEA case studies to identify the extent of participation, learning outcomes attributable to participation, and if any learning outcomes led to social action for sustainability at the community level. Strengths across the two cases were the inclusion of marginalized populations and consideration of socio-economic concerns. Consistent weaknesses included inadequate notice, document inaccessibility, lack of feedback and communication, and late analysis of alternatives. Despite some learning conditions being unfulfilled, examples of instrumental, communicative, and transformative learning were identified through a focus group and semi-structured interviews with community participants and public officials. Some of these learning outcomes led to individual and social actions that contribute to sustainability. -- Highlights: • The strengths and weaknesses of Kenyan SEA public participation processes were identified. • Multiple deficiencies in the SEA process likely frustrate meaningful public engagement. • Participant learning was observed despite process weaknesses. • Participant learning can lead to action for sustainability at the community level.

54 ENVIRONMENTAL SCIENCES↗

Backpropagation-based learning with local derivative approximation and memory replay in biologically plausible neural systems

When learning, the brain modifies individual synaptic connections to reach a desired behavior. Animal and human brains have been shown to be incredibly capable of learning complex and varied functions across a wide variety of tasks. In recent years, artificial neural networks, inspired by human and animal brains, have shown great capabilities in learning a wide variety of difficult tasks. However, artificial neural networks primarily teach themselves through the use of backpropagation, a learning method which has no clear analogue within the brain. Additionally, Artificial Neural Networks primarily use continuous activation functions, which differ significantly from the spiking neuronal behavior present in the brain. In this paper, we discuss and demonstrate a biologically plausible learning method that approximates backpropagation through two techniques on Spiking Neural Networks. First, we show that the local temporal derivatives that are necessary for backpropagation can be approximately recovered through reconstruction using spike timings. Second, we show that through learning during a sleep phase, inspired by neuroscience research into memory replay, the localized parallel feedback path can learn to approximate the derivative through the forward path weight matrix, thus solving the weight transport problem. Lastly, we demonstrate that the combination of these two methods can approach or exceed the accuracy of backpropagation-based methods for a variety of neuromorphic vision tasks while maintaining biological plausibility.

42 ENGINEERING↗

Ensemble transfer learning for the prediction of anti-cancer drug response

Abstract Transfer learning, which transfers patterns learned on a source dataset to a related target dataset for constructing prediction models, has been shown effective in many applications. In this paper, we investigate whether transfer learning can be used to improve the performance of anti-cancer drug response prediction models. Previous transfer learning studies for drug response prediction focused on building models to predict the response of tumor cells to a specific drug treatment. We target the more challenging task of building general prediction models that can make predictions for both new tumor cells and new drugs. Uniquely, we investigate the power of transfer learning for three drug response prediction applications including drug repurposing, precision oncology, and new drug development, through different data partition schemes in cross-validation. We extend the classic transfer learning framework through ensemble and demonstrate its general utility with three representative prediction algorithms including a gradient boosting model and two deep neural networks. The ensemble transfer learning framework is tested on benchmark in vitro drug screening datasets. The results demonstrate that our framework broadly improves the prediction performance in all three drug response prediction applications with all three prediction algorithms.

60 APPLIED LIFE SCIENCES↗

Prospects of federated machine learning in fluid dynamics

Physics-based models have been mainstream in fluid dynamics for developing predictive models. In recent years, machine learning has offered a renaissance to the fluid community due to the rapid developments in data science, processing units, neural network based technologies, and sensor adaptations. So far in many applications in fluid dynamics, machine learning approaches have been mostly focused on a standard process that requires centralizing the training data on a designated machine or in a data center. In this article, we present a federated machine learning approach that enables localized clients to collaboratively learn an aggregated and shared predictive model while keeping all the training data on each edge device. We demonstrate the feasibility and prospects of such a decentralized learning approach with an effort to forge a deep learning surrogate model for reconstructing spatiotemporal fields. Our results indicate that federated machine learning might be a viable tool for designing highly accurate predictive decentralized digital twins relevant to fluid dynamics.

36 MATERIALS SCIENCE↗

MO-E-18C-03: Incorporating Active Learning Into A Traditional Graduate Medical Physics Course

Purpose: To improve the ability of graduate students to learn medical physics concepts through the incorporation of active learning techniques. Methods: A traditional lecture-based radiological physics course was modified such that: (1) traditional (two-hour) lectures were provided online for students to watch prior to class, (2) a student was chosen randomly at the start of each class to give a two minute synopsis of the material and its relevance (two-minute drill), (3) lectures were significantly abbreviated and remaining classroom time used for group problem solving, and (4) videos of the abbreviated lectures were made available online for review. In the transition year, students were surveyed about the perceived effects of these changes on learning. Student performance was evaluated for 3 years prior to and 4 years after modification. Results: The survey tool used a five point scale from 1=Not True to 5=Very True. While nearly all students reviewed written materials prior to class (4.3±0.9), a minority watched the lectures (2.1±1.5). A larger number watched the abbreviated lectures for further clarification (3.6±1.6) and found it helpful in learning the content (4.2±1.0). Most felt that the two-minute drill helped them get more out of the lecture (3.9±0.8) and the problem solving contributed to their understanding of the content (4.1±0.8). However, no significant improvement in exam scores resulted from the modifications (mean scores well within 1 SD during study period). Conclusion: Students felt that active learning techniques improved their ability to learn the material in what is considered the most difficult course in the program. They valued the ability to review the abbreviated class lecture more than the opportunity to watch traditional lectures prior to class. While no significant changes in student performance were observed, aptitude variations across the student cohorts make it difficult to draw conclusions about the effectiveness of active learning.

96 KNOWLEDGE MANAGEMENT AND PRESERVATION↗

SU-F-P-20: Predicting Waiting Times in Radiation Oncology Using Machine Learning

Purpose: Waiting times remain one of the most vexing patient satisfaction challenges facing healthcare. Waiting time uncertainty can cause patients, who are already sick or in pain, to worry about when they will receive the care they need. These waiting periods are often difficult for staff to predict and only rough estimates are typically provided based on personal experience. This level of uncertainty leaves most patients unable to plan their calendar, making the waiting experience uncomfortable, even painful. In the present era of electronic health records (EHRs), waiting times need not be so uncertain. Extensive EHRs provide unprecedented amounts of data that can statistically cluster towards representative values when appropriate patient cohorts are selected. Predictive modelling, such as machine learning, is a powerful approach that benefits from large, potentially complex, datasets. The essence of machine learning is to predict future outcomes by learning from previous experience. The application of a machine learning algorithm to waiting time data has the potential to produce personalized waiting time predictions such that the uncertainty may be removed from the patient’s waiting experience. Methods: In radiation oncology, patients typically experience several types of waiting (eg waiting at home for treatment planning, waiting in the waiting room for oncologist appointments and daily waiting in the waiting room for radiotherapy treatments). A daily treatment wait time model is discussed in this report. To develop a prediction model using our large dataset (with more than 100k sample points) a variety of machine learning algorithms from the Python package sklearn were tested. Results: We found that the Random Forest Regressor model provides the best predictions for daily radiotherapy treatment waiting times. Using this model, we achieved a median residual (actual value minus predicted value) of 0.25 minutes and a standard deviation residual of 6.5 minutes. This means that the majority of our estimates are within 6.5 minutes of the actual wait time. Conclusion: The goal of this project was to define an appropriate machine learning algorithm to estimate waiting times based on the collective knowledge and experience learned from previous patients. Our results offer an opportunity to improve the information that is provided to patients and family members regarding the amount of time they can expect to wait for radiotherapy treatment at our centre. AJ acknowledges support by the CREATE Medical Physics Research Training Network grant of the Natural Sciences and Engineering Research Council (Grant number: 432290) and from the 2014 Q+ Initiative of the McGill University Health Centre.

60 APPLIED LIFE SCIENCES↗

TH-CD-206-05: Machine-Learning Based Segmentation of Organs at Risks for Head and Neck Radiotherapy Planning

Purpose: Accurate and efficient delineation of tumor target and organs-at-risks is essential for the success of radiotherapy. In reality, despite of decades of intense research efforts, auto-segmentation has not yet become clinical practice. In this study, we present, for the first time, a deep learning-based classification algorithm for autonomous segmentation in head and neck (HaN) treatment planning. Methods: Fifteen HN datasets of CT, MR and PET images with manual annotation of organs-at-risk (OARs) including spinal cord, brainstem, optic nerves, chiasm, eyes, mandible, tongue, parotid glands were collected and saved in a library of plans. We also have ten super-resolution MR images of the tongue area, where the genioglossus and inferior longitudinalis tongue muscles are defined as organs of interest. We applied the concepts of random forest- and deep learning-based object classification for automated image annotation with the aim of using machine learning to facilitate head and neck radiotherapy planning process. In this new paradigm of segmentation, random forests were used for landmark-assisted segmentation of super-resolution MR images. Alternatively to auto-segmentation with random forest-based landmark detection, deep convolutional neural networks were developed for voxel-wise segmentation of OARs in single and multi-modal images. The network consisted of three pairs of convolution and pooing layer, one RuLU layer and a softmax layer. Results: We present a comprehensive study on using machine learning concepts for auto-segmentation of OARs and tongue muscles for the HaN radiotherapy planning. An accuracy of 81.8% in terms of Dice coefficient was achieved for segmentation of genioglossus and inferior longitudinalis tongue muscles. Preliminary results of OARs regimentation also indicate that deep-learning afforded an unprecedented opportunities to improve the accuracy and robustness of radiotherapy planning. Conclusion: A novel machine learning framework has been developed for image annotation and structure segmentation. Our results indicate the great potential of deep learning in radiotherapy treatment planning.

60 APPLIED LIFE SCIENCES↗

Data and scripts associated with the manuscript "Encoding Diel Hysteresis and the Birch Effect in Dryland Soil Respiration Models through Knowledge-Guided Deep Learning"

This package contains the data and scripts used in "Encoding Diel Hysteresis and the Birch Effect in Dryland Soil Respiration Models through Knowledge-Guided Deep Learning" (Jiang et al., 2022). The data.zip file contains the flux tower and automated chamber observations used for developing the deep learning model for modeling soil respiration. The scripts.zip file contains the Jupyter notebooks and python scripts for preprocessing the data, training the deep learning models, and postprocessing the results. The src.zip contains the source code for training the deep learning model, performing mutual information analysis, and plotting functions. The trained_models.zip contains multiple folders used for hosting the trained deep-learning models and the associated soil respiration predictions. The whole process is performed using python. We include the REAMD.md to document the python package requirements.Soil respiration in dryland ecosystems is challenging to model due to its complex interactions with environmental drivers. Knowledge-guided deep learning provides a much more effective means of accurately representing these complex interactions than traditional Q10-based models. Mutual information analysis revealed that future soil temperature shares more information with soil respiration than past soil temperature, consistent with their clockwise diel hysteresis. We explicitly encoded diel hysteresis, soil drying, and soil rewetting effects on soil respiration dynamics in a newly designed Long Short Term Memory (LSTM) model. The model takes both past and future environmental drivers as inputs to predict soil respiration. The new LSTM model substantially outperformed three Q10-based models and the Community Land Model when reproducing the observed soil respiration dynamics in a semi-arid ecosystem. The new LSTM model clearly demonstrated its superiority for temporally extrapolating soil respiration dynamics, such that the resulting correlation with observational data is up to 0.7 while the correlations of both Q10-based models and the Community Land Model (CLM) are less than 0.4. Our results underscore the high potential for knowledge-guided deep learning to replace Q10-based soil respiration modules in Earth system models.

54 ENVIRONMENTAL SCIENCES↗

LENS: Learning Enabled Network Synthesis

RTRC and UMD have developed novel machine learning based methods under the ARPA-E DIFFERENTIATE program for rapid acceleration of hypothesis generation in complex architecture design spaces involving both discrete choices of component inclusion and interconnection and continuous parametric decisions. The project named Learning Enabled Network Synthesis (LENS) further demonstrated the developed methods on challenging electrical power converter design problems by identifying the most suitable circuit topologies and simultaneously selecting the most appropriate components to achieve optimized design of power converter with improved performances. We demonstrated that LENS could enable exploration of very large design space of circuit topologies and components by addressing the limitations of conventional design process in non-linear, high switching speed, multi-dimensional power converter design and optimization. The key innovation developed in LENS is the seamless integration of statistical learning and logical reasoning techniques and building on the individual strengths of these techniques for rapid hypothesis discovery. The main component of LENS comprises of: 1) Graph Reasoning Engine (GRE) to enforce composition rules that rapidly reject all discrete architectures that are composed incorrectly and generates an adaptive database of feasible designs which can be used by ML modules, 2) Graph Generative Learning module which is a deep neural network based generative model for graph architectures which can enable design space exploration beyond the dataset generated by the GRE, 3) Graph Reduced Order Model (ROM) for graph domains for accelerating computation of output metrics, and 4) Active learning and Rule Discovery module for sample efficient learning and extracting logical rules from the learned ML models which will be integrated in the GRE to enhance the filtering effectiveness. LENS approach can be applied to any design domains where designs can be represented as multi-attribute graphs. The LENS team integrated the various technical innovations listed above into an optimization pipeline and exercised the optimization pipeline on the converter design problem. The LENS project demonstrated that the developed AI/ML technologies can be used to generate novel converter circuits >45x faster than experts on chosen use-cases. This can enable faster design space exploration and identification of new designs which are not considered by experts due to the increasing design space complexity. This has significant potential impact on the public and energy needs of the country. It is currently estimated that 30% of all electrical powers generated passes through power converters. The future estimate is that 80% of all power generated would be passing through converters. LENS fills a critical gap in this space since by accelerating the design process the designers would be able to generate more efficient converters which can lead to significant energy savings for the country.

42 ENGINEERING↗

Using Machine Learning to Predict Core Sizes of High-Efficiency Turbofan Engines

With the rise in big data and analytics, machine learning is transforming many industries. It is being increasingly employed to solve a wide range of complex problems, producing autonomous systems that support human decision-making. For the aircraft engine industry, machine learning of historical and existing engine data could provide insights that help drive for better engine design. This work explored the application of machine learning to engine preliminary design. Engine core-size prediction was chosen for the first study because of its relative simplicity in terms of number of input variables required (only three). Specifically, machine-learning predictive tools were developed for turbofan engine core-size prediction, using publicly available data of two hundred manufactured engines and engines that were studied previously in NASA aeronautics projects. The prediction results of these models show that, by bringing together big data, robust machine-learning algorithms and data science, a machine learning-based predictive model can be an effective tool for turbofan engine core-size prediction. The promising results of this first study paves the way for further exploration of the use of machine learning for aircraft engine preliminary design.

Core Size↗

Aero-Engines AI - A Machine-Learning App for Aircraft Engine Concepts Assessment

Effective deployment of trained machine-learning models could drive a high level of efficiency in aircraft engine conceptual design. Aero-Engines AI is a Windows app that has been created to deploy trained machine-learning models to assess aircraft engine concepts. It was created using tkinter, a GUI (graphical user interface) module that is built into the standard Python library. Employing tkinter greatly facilitates the sharing of machine-learning application as an executable file which can be run on Windows machines (without the need to have Python or any library installed). Current version of the app focuses on the performance prediction of conventional turbofans. The app gets user input for a turbofan design, preprocesses the input data, and deploys trained machine-learning models to predict turbofan thrust specific fuel consumption (TSFC), engine weight, core size, and turbomachinery stage-counts. The machine-learning predictive models were built by employing supervised deep-learning algorithm to study patterns in an existing open-source database of production and research turbofan engines. They were trained, cross-validated, and tested in Keras, an open-source neural networks API (application programming interface) written in Python, with TensorFlow (Google open-source artificial intelligence library) serving as the backend engine. The smooth deployment of these machine-learning models using the app shows that Aero-Engines AI is an easy-to-use and a time-saving tool for aircraft engine design-space exploration during the conceptual design stage.

machine learning↗

Machine Learning in the Big Data Era: Are We There Yet?

In this paper, we discuss the machine learning challenges of the Big Data era. We observe that recent innovations in being able to collect, access, organize, integrate, and query massive amounts of data from a wide variety of data sources have brought statistical machine learning under more scrutiny and evaluation for gleaning insights from the data than ever before. In that context, we pose and debate the question - Are machine learning algorithms scaling with the ability to store and compute? If yes, how? If not, why not? We survey recent developments in the state-of-the-art to discuss emerging and outstanding challenges in the design and implementation of machine learning algorithms at scale. We leverage experience from real-world Big Data knowledge discovery projects across domains of national security and healthcare to suggest our efforts be focused along the following axes: (i) the data science challenge - designing scalable and flexible computational architectures for machine learning (beyond just data-retrieval); (ii) the science of data challenge the ability to understand characteristics of data before applying machine learning algorithms and tools; and (iii) the scalable predictive functions challenge the ability to construct, learn and infer with increasing sample size, dimensionality, and categories of labels. We conclude with a discussion of opportunities and directions for future research.

Big Data↗

Scientific machine learning for closure models in multiscale problems: A review

Here, closure problems are omnipresent when simulating multiscale systems, where some quantities and processes cannot be fully prescribed despite their effects on the simulation's accuracy. Recently, scientific machine learning approaches have been proposed as a way to tackle the closure problem, combining traditional (physics-based) modeling with data-driven (machine-learned) techniques, typically through enriching differential equations with neural networks. This paper reviews the different reduced model forms, distinguished by the degree to which they include known physics, and the different objectives of a priori and a posteriori learning. The importance of adhering to physical laws (such as symmetries and conservation laws) in choosing the reduced model form and choosing the learning method is discussed. The effect of spatial and temporal discretization and recent trends toward discretization-invariant models are reviewed. In addition, we make the connections between closure problems and several other research disciplines: inverse problems, Mori-Zwanzig theory, and multi-fidelity methods. In conclusion, much progress has been made with scientific machine learning approaches for solving closure problems, but many challenges remain. In particular, the generalizability and interpretability of learned models is a major issue that needs to be addressed further.

97 MATHEMATICS AND COMPUTING↗

Machine learning materials physics: Surrogate optimization and multi-fidelity algorithms predict precipitate morphology in an alternative to phase field dynamics

Machine learning has been effective at detecting patterns and predicting the response of systems that behave free of natural laws. Examples include learning crowd dynamics, recommender systems and autonomous mobility. There also have been applications to the search for new materials that draw upon big-data classification problems. Though, when it comes to physical systems governed by conservation laws, the role of machine learning has been more limited. Here, we present our recent work in exploring the role of machine learning methods in discovering, or aiding, the search for physics. Specifically, we focus on using machine learning algorithms to represent high-dimensional free energy surfaces with the goal of identifying precipitate morphologies in alloy systems. Traditionally, this problem has been approached by combining phase field models, which impose first-order dynamics, with elasticity, to traverse a free energy landscape in search of minima. Equilibrium precipitate morphologies occur at these minima. Here, we exploit the machine learning methods to represent high-dimensional data, combined with surrogate optimization, sensitivity analysis and multifidelity modeling as an alternate framework to explore phenomena controlled by energy extremization. This combination of data-driven methods offers an alternative to the imposition of first-order dynamics via phase field methods, and represents one aspect of applying machine learning to materials physics.

42 ENGINEERING↗

Assessing Adaptive Irrigation Impacts on Water Scarcity in Nonstationary Environments—A Multi‐Agent Reinforcement Learning Approach

Abstract One major challenge in water resource management is to balance the uncertain and nonstationary water demands and supplies caused by the changing anthropogenic and hydroclimate conditions. To address this issue, we developed a reinforcement learning agent‐based modeling (RL‐ABM) framework where agents (agriculture water users) are able to learn and adjust water demands based on their interactions with the water systems. The intelligent agents are created by a reinforcement learning algorithm adapted from the Q‐learning algorithm. We illustrated this framework in a case study where the RL‐ABM is two‐way coupled with the Colorado River Simulation System (CRSS), a long‐term planning model used for the administration of the Colorado River Basin, for assessing agriculture water uses impacts on water scarcity. Seventy‐eight intelligent agents are simulated, which can be grouped into three categories based on their parameter values: the “aggressive” (swift actions; low regrets), the “forward‐looking conservative” (mild actions; high regrets; fast learning), and the “myopic conservative” (mild actions; median regrets; slow learning). The ABM‐CRSS results showed that the major reservoirs in the Upper Colorado Basin might experience more frequent water shortages due to the increasing water uses compared to the original CRSS results. If the drought continues, the case study also demonstrates that agents can learn and adjust their demands.

Hung, Fengwei↗

Causally‐Informed Deep Learning to Improve Climate Models and Projections

Abstract Climate models are essential to understand and project climate change, yet long‐standing biases and uncertainties in their projections remain. This is largely associated with the representation of subgrid‐scale processes, particularly clouds and convection. Deep learning can learn these subgrid‐scale processes from computationally expensive storm‐resolving models while retaining many features at a fraction of computational cost. Yet, climate simulations with embedded neural network parameterizations are still challenging and highly depend on the deep learning solution. This is likely associated with spurious non‐physical correlations learned by the neural networks due to the complexity of the physical dynamical system. Here, we show that the combination of causality with deep learning helps removing spurious correlations and optimizing the neural network algorithm. To resolve this, we apply a causal discovery method to unveil causal drivers in the set of input predictors of atmospheric subgrid‐scale processes of a superparameterized climate model in which deep convection is explicitly resolved. The resulting causally‐informed neural networks are coupled to the climate model, hence, replacing the superparameterization and radiation scheme. We show that the climate simulations with causally‐informed neural network parameterizations retain many convection‐related properties and accurately generate the climate of the original high‐resolution climate model, while retaining similar generalization capabilities to unseen climates compared to the non‐causal approach. The combination of causal discovery and deep learning is a new and promising approach that leads to stable and more trustworthy climate simulations and paves the way toward more physically‐based causal deep learning approaches also in other scientific disciplines.

Meteorology & Atmospheric Sciences↗