NashAE: Disentangling Representations through Adversarial Covariance Minimization
Explore the source record for details and available documents.
Yeats, Eric↗
SEARCH · Engineering Papers
Results for “disentangled representations”
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.
Optimizing training trajectories in variational autoencoders via latent Bayesian optimization approach *
Unsupervised and semi-supervised ML methods such as variational autoencoders (VAE) have become widely adopted across multiple areas of physics, chemistry, and materials sciences due to their capability in disentangling representations and ability to find latent manifolds for classification and/or regression of complex experimental data. Like other ML problems, VAEs require hyperparameter tuning, e.g. balancing the Kullback–Leibler and reconstruction terms. However, the training process and resulting manifold topology and connectivity depend not only on hyperparameters, but also their evolution during training. Because of the inefficiency of exhaustive search in a high-dimensional hyperparameter space for the expensive-to-train models, here we have explored a latent Bayesian optimization (zBO) approach for the hyperparameter trajectory optimization for the unsupervised and semi-supervised ML and demonstrated for joint-VAE with rotational invariances. We have demonstrated an application of this method for finding joint discrete and continuous rotationally invariant representations for modified national institute of standards and technology database (MNIST) and experimental data of a plasmonic nanoparticles material system. The performance of the proposed approach has been discussed extensively, where it allows for any high dimensional hyperparameter trajectory optimization of other ML models.
42 ENGINEERING↗
Machine Learning Explainability and Transferability for Path Navigation
Deep neural networks are powerful tools for machine perception. Unfortunately their decisions are difficult to explain due to the complexity and size of the networks. Previously we have alleviated this issue by using the representational portion of a deep neural network and combining it with a k-nearest neighbor (KNN) classifier. Through inspection of the decisions made by the KNN, we can directly see the training data responsible for the decisions, allowing us to determine the quality of the overall decision and the quality of the representational layer of the deep NN. While the technique worked well, it requires tens of thousands of latent vectors to be stored for classification. In addition, it lacks the ability to show how parts of an image influence the classification decision. Here we address these issues by 1) Using a radial basis function network (RBFN) in place of the KNN allowing far fewer images to be used in deployment and 2) Using an auto encoder network for explainability. In addition to these techniques, we examine the effects of transfer learning to determine that results are robust. All results are tested on a domain where an unmanned aerial vehicle (UAV) navigates a forest trail through a single camera.
explainable machine learning↗
Uncovering obscured phonon dynamics from powder inelastic neutron scattering using machine learning
The study of phonon dynamics is pivotal for understanding material properties, yet it faces challenges due to the irreversible information loss inherent in powder inelastic neutron scattering spectra and the limitations of traditional analysis methods. In this study, we present a machine learning framework designed to reveal obscured phonon dynamics from powder spectra. Using a variational autoencoder, we obtain a disentangled latent representation of spectra and successfully extract force constants for reconstructing phonon dispersions. Notably, our model demonstrates effective applicability to experimental data even when trained exclusively on physics-based simulations. The fine-tuning with experimental spectra further mitigates issues arising from domain shift. Analysis of latent space underscores the model’s versatility and generalizability, affirming its suitability for complex system applications. Furthermore, our framework’s two-stage design is promising for developing a universal pre-trained feature extractor. This approach has the potential to revolutionize neutron measurements of phonon dynamics, offering researchers a potent tool to decipher intricate spectra and gain valuable insights into the intrinsic physics of materials.
domain adaptation↗
Disentangling Rotational Dynamics and Ordering Transitions in a System of Self-Organizing Protein Nanorods via Rotationally Invariant Latent Representations
The dynamics of complex ordering systems with active rotational degrees of freedom exemplified by protein self-assembly is explored using a machine learning workflow that combines deep learning-based semantic segmentation and rotationally invariant variational autoencoder-based analysis of orientation and shape evolution. The latter allows for disentanglement of the particle orientation from other degrees of freedom and compensates for lateral shifts. The disentangled representations in the latent space encode the rich spectrum of local transitions that can now be visualized and explored via continuous variables. The time dependence of ensemble averages allows insight into the time dynamics of the system and, in particular, illustrates the presence of the potential ordering transition. Finally, analysis of the latent variables along the single-particle trajectory allows tracing these parameters on a single-particle level. The proposed approach is expected to be universally applicable for the description of the imaging data in optical, scanning probe, and electron microscopy seeking to understand the dynamics of complex systems where rotations are a significant part of the process.
representation learning↗
A Representation Fusion Framework for Decoupling Diagnostic Information in Multimodal Learning
Modern medicine increasingly relies on multimodal data, ranging from clinical notes to imaging and genomics, to guide diagnosis and treatment. However, integrating these heterogeneous data sources in a principled and interpretable manner remains a major challenge. We present MODES (Multi-mOdal Disentangled Embedding Space), a representation fusion framework that explicitly separates shared and modality-specific factors of variation, offering a structured latent space for multimodal information that improves both prediction and interpretability. By leveraging pre-trained unimodal foundation models, MODES mitigates the dependency on extensive paired datasets, crucial in data-scarce clinical settings. We introduce a masking strategy that optimizes representation dimensionality by eliminating low-information dimensions, to achieve compact, information-rich representations. Our framework demonstrates superior performance in predicting diagnoses and phenotypes compared to unimodal and conventional fusion models. MODES also enables robust diagnostic inference in missing data scenarios, offering an opportunity toward interpretable and efficient multimodal diagnostics in personalized healthcare.
60 APPLIED LIFE SCIENCES↗
Soil Moisture‐Cloud‐Precipitation Feedback in the Lower Atmosphere From Functional Decomposition of Satellite Observations
Abstract The feedback of topsoil moisture (SM) content on convective clouds and precipitation is not well understood and represented in the current generation of weather and climate models. Here, we use functional decomposition of satellite‐derived SM and cloud vertical profiles (CVP) to quantify the relationship between SM and the vertical distribution of cloud water in the central US. High‐dimensional model representation is used to disentangle the contributions of SM and other land‐surface and atmospheric variables to the CVP. Results show that the sign and strength of the SM‐cloud‐precipitation feedback varies with cloud height and time lag and displays a large spatial variability. Positive anomalies in antecedent 7‐hr SM and land‐surface temperature enhance cloud reflectivity up to 4 dBZ in the lower atmosphere about 1–3 km above the surface. Our approach presents new insights into the SM‐cloud‐precipitation feedback and aids in the diagnosis of land‐atmosphere interactions simulated by weather and climate models.
54 ENVIRONMENTAL SCIENCES↗
CAD-VAE: Leveraging Correlation-Aware Latents for Comprehensive Fair Disentanglement
While deep generative models have significantly advanced representation learning, they may inherit or amplify biases and fairness issues by encoding sensitive attributes alongside predictive features. Enforcing strict independence in disentanglement is often unrealistic when target and sensitive factors are naturally correlated. To address this challenge, we propose CAD-VAE(Correlation-Aware Disentangled VAE), which introduces a correlated latent code to capture the information shared between the target and sensitive attributes. Given this correlated latent, our method effectively separates over-lapping factors without extra domain knowledge by directly minimizing the conditional mutual information between target and sensitive codes. A relevance-driven optimization strategy refines the correlated code by efficiently capturing essential correlated features and eliminating redundancy. Extensive experiments on benchmark datasets demonstrate that CAD-VAE produces fairer representations, realistic counterfactuals, and improved fairness-aware image editing.
Ma, Chenrui [University of California Irvine]↗
Sensitivity of aerosol optical properties to the aerosol size distribution over central Europe and the Mediterranean Basin using the WRF-Chem v.3.9.1.1 coupled model
The size distribution of atmospheric aerosols plays a key role for understanding and quantifying the uncertainties related to aerosol–radiation and aerosol–cloud interactions. These interactions ultimately depend on the size distribution through optical properties (such as aerosol optical depth, AOD) or cloud microphysical properties. Hence, the main objective of this contribution is to disentangle the impact of the representation of aerosol size distribution on aerosol optical properties over central Europe, particularly over the Mediterranean Basin, during a summertime aerosol episode. To fulfill this objective, a sensitivity test has been conducted using the coupled chemistry–meteorology model WRF-Chem (Weather Research Forecast model coupled with Chemistry). The test modified the parameters defining a lognormal size distribution (geometric diameter and standard deviation) by 10 %, 20 %, and 50 %. Results reveal that the reduction in the standard deviation of the accumulation mode leads to the largest impacts on AOD due to a transfer of particles from the accumulation mode to the coarse mode. A reduction in the geometric diameter of the accumulation mode also has an influence on AOD representation since particles in this mode are assumed to be smaller. In addition, an increase in the geometric diameter of the coarse mode produces a redistribution through the total size distribution by relocating particles from the finer modes to the coarse.
54 ENVIRONMENTAL SCIENCES↗
A robust estimator of mutual information for deep learning interpretability
Abstract We develop the use of mutual information (MI), a well-established metric in information theory, to interpret the inner workings of deep learning (DL) models. To accurately estimate MI from a finite number of samples, we present GMM-MI (pronounced ‘Jimmie’), an algorithm based on Gaussian mixture models that can be applied to both discrete and continuous settings. GMM-MI is computationally efficient, robust to the choice of hyperparameters and provides the uncertainty on the MI estimate due to the finite sample size. We extensively validate GMM-MI on toy data for which the ground truth MI is known, comparing its performance against established MI estimators. We then demonstrate the use of our MI estimator in the context of representation learning, working with synthetic data and physical datasets describing highly non-linear processes. We train DL models to encode high-dimensional data within a meaningful compressed (latent) representation, and use GMM-MI to quantify both the level of disentanglement between the latent variables, and their association with relevant physical quantities, thus unlocking the interpretability of the latent representation. We make GMM-MI publicly available in this GitHub repository.
46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
Physics and chemistry from parsimonious representations: image analysis via invariant variational autoencoders
Electron, optical, and scanning probe microscopy methods are generating ever increasing volume of image data containing information on atomic and mesoscale structures and functionalities. This necessitates the development of the machine learning methods for discovery of physical and chemical phenomena from the data, such as manifestations of symmetry breaking phenomena in electron and scanning tunneling microscopy images, or variability of the nanoparticles. Variational autoencoders (VAEs) are emerging as a powerful paradigm for the unsupervised data analysis, allowing to disentangle the factors of variability and discover optimal parsimonious representation. Here, we summarize recent developments in VAEs, covering the basic principles and intuition behind the VAEs. The invariant VAEs are introduced as an approach to accommodate scale and translation invariances present in imaging data and separate known factors of variations from the ones to be discovered. We further describe the opportunities enabled by the control over VAE architecture, including conditional, semi-supervised, and joint VAEs. Several case studies of VAE applications for toy models and experimental datasets in Scanning Transmission Electron Microscopy are discussed, emphasizing the deep connection between VAE and basic physical principles. Python codes and datasets discussed in this article are available at https://github.com/saimani5/VAE-tutorials and can be used by researchers as an application guide when applying these to their own datasets.
36 MATERIALS SCIENCE↗
Enhancing Interpretability in Generative Modeling: Statistically Disentangled Latent Spaces Guided by Generative Factors in Scientific Datasets
This study addresses the challenge of statistically extracting generative factors from complex, high-dimensional datasets in unsupervised or semi-supervised settings. We investigate encoder-decoder-based generative models for nonlinear dimensionality reduction, focusing on disentangling low-dimensional latent variables corresponding to independent physical factors. Introducing Aux-VAE, a novel architecture within the classical Variational Autoencoder framework, we achieve disentanglement with minimal modifications to the standard VAE loss function by leveraging prior statistical knowledge through auxiliary variables. These variables guide the shaping of the latent space by aligning latent factors with learned auxiliary variables. We validate the efficacy of Aux-VAE through comparative assessments on multiple datasets, including astronomical simulations.
97 MATHEMATICS AND COMPUTING↗
Probing atomic-scale symmetry breaking by rotationally invariant machine learning of multidimensional electron scattering
Abstract The 4D scanning transmission electron microscopy (STEM) method maps the structure and functionality of solids on the atomic scale, yielding information-rich data sets describing the interatomic electric and magnetic fields, structural and electronic order parameters, and other symmetry breaking distortions. A critical bottleneck is the dearth of analytical tools that can reduce complex 4D-STEM data to physically relevant descriptors. We propose an approach for the systematic exploration of 4D-STEM data using rotationally invariant variational autoencoders (rrVAE), which disentangle the general rotation of the object from other latent representations. The implementation of purely rotational rrVAE is discussed as are applications to simulated data for graphene and zincblende structures. The rrVAE analysis of experimental 4D-STEM data of defects in graphene is illustrated and compared to the classical center-of-mass analysis. This approach is universal for probing symmetry-breaking phenomena in complex systems and can be implemented for a broad range of diffraction methods.
36 MATERIALS SCIENCE↗
Interactively Assessing Disentanglement in GANs
Abstract Generative adversarial networks (GAN) have witnessed tremendous growth in recent years, demonstrating wide applicability in many domains. However, GANs remain notoriously difficult for people to interpret, particularly for modern GANs capable of generating photo‐realistic imagery. In this work we contribute a visual analytics approach for GAN interpretability, where we focus on the analysis and visualization of GAN disentanglement. Disentanglement is concerned with the ability to control content produced by a GAN along a small number of distinct, yet semantic, factors of variation. The goal of our approach is to shed insight on GAN disentanglement, above and beyond coarse summaries, instead permitting a deeper analysis of the data distribution modeled by a GAN. Our visualization allows one to assess a single factor of variation in terms of groupings and trends in the data distribution, where our analysis seeks to relate the learned representation space of GANs with attribute‐based semantic scoring of images produced by GANs. Through use‐cases, we show that our visualization is effective in assessing disentanglement, allowing one to quickly recognize a factor of variation and its overall quality. In addition, we show how our approach can highlight potential dataset biases learned by GANs.
Jeong, Sangwon↗
Causal Interactions between Southern Ocean Polynyas and High-Latitude Atmosphere–Ocean Variability
Abstract Weddell Sea open-ocean polynyas have been observed to occasionally release heat from the deep ocean to the atmosphere, indicating that their sporadic appearances may be an important feature of high-latitude atmosphere–ocean variability. Yet, observations of the phenomenon are sparse and many standard-resolution models represent these features poorly, if at all. We use a fully coupled, synoptic-scale preindustrial control simulation of the Energy Exascale Earth System Model (E3SMv0-HR) to effectively simulate open-ocean polynyas and investigate their role in the climate system. Our approach employs statistical tests of Granger causality to diagnose local and remote drivers of, and responses to, polynya heat loss on interannual to decadal time scales. First, we find that polynya heat loss Granger causes a persistent increase in surface air temperature over the Weddell Sea, strengthening the local cyclonic wind circulation. Along with responding to polynyas, atmospheric conditions also facilitate their development. When the Southern Ocean experiences a rapid poleward shift in the circumpolar westerlies following a prolonged negative phase of the southern annular mode (SAM), Weddell Sea salinity increases, promoting density destratification and convection in the water column. Finally, we find that the reduction of surface heat fluxes during periods of full ice cover is not fully compensated by ocean heat transport into the high latitudes. This imbalance leads to a buildup of ocean heat content that supplies polynya heat loss. These results disentangle the complex, coupled climate processes that both enable the polynya’s existence and respond to it, providing insights to improve the representation of these highly episodic sea ice features in climate models.
Kaufman, Zachary S.↗
NuGraph2 with context-aware inputs: physics-inspired improvements in semantic segmentation
Graph neural networks have recently shown strong promise for event reconstruction tasks in Liquid Argon Time Projection Chambers, yet their performance remains limited for underrepresented classes of particles, such as Michel electrons. In this work, we investigate physics-informed strategies to improve semantic segmentation within the NuGraph2 architecture. We explore three complementary approaches: (i) enriching the input representation with context-aware features derived from detector geometry and track continuity, (ii) introducing auxiliary decoders to capture class-level correlations, and (iii) incorporating energy-based regularization terms motivated by Michel electron energy distributions. Experiments on MicroBooNE public datasets show that physics-inspired feature augmentation yields the largest gains, particularly boosting Michel electron precision and recall by disentangling overlapping latent space regions. In contrast, auxiliary decoders and energy-regularization terms provided limited improvements, partly due to the hit-level nature of NuGraph2, which lacks explicit particle- or event-level representations. Our findings highlight that embedding physics context directly into node-level inputs is more effective than imposing task-specific auxiliary losses, and suggest that future hierarchical architectures such as NuGraph3, with explicit particle- and event-level reasoning, will provide a more natural setting for advanced decoders and physics-based regularization. The code for this work is publicly available on Github at https://github.com/vitorgrizzi/nugraph_phys/tree/main_phys.
Other Experiments↗
Disentangling Ferroelectric Wall Dynamics and Identification of Pinning Mechanisms via Deep Learning
Field-induced domain-wall dynamics in ferroelectric materials underpins multiple applications ranging from actuators to information technology devices and necessitates a quantitative description of the associated mechanisms including giant electromechanical couplings, controlled nonlinearities, or low coercive voltages. While the advances in dynamic piezoresponse force microscopy measurements over the last two decades have rendered visualization of polarization dynamics relatively straightforward, the associated insights into the local mechanisms have been elusive. This work explores the domain dynamics in model polycrystalline materials using a workflow combining deep-learning-based segmentation of the domain structures with nonlinear dimensionality reduction using multilayer rotationally invariant autoencoders (rVAE). The former allows unambiguous identification and classification of the ferroelectric and ferroelastic domain walls. Here, the rVAE discovers the latent representations of the domain wall geometries and their dynamics, consequently providing insight into the intrinsic mechanisms of polarization switching, that can further be compared to simple physical models. The rVAE disentangles the factors affecting the pinning efficiency of ferroelectric walls, offering insights into the correlation of ferroelastic wall distribution and ferroelectric wall pinning.
36 MATERIALS SCIENCE↗
Disentangling the confounding spectroscopy of C 1 molecules: Without symmetry as a guide, everything is allowed
The spectra of C 1 molecules are confounding in that each of the fundamental vibrational modes transform as the same irreducible representation (A) and hence each band consists of a seemingly random distribution of a-, b-, and c-type transitions. This is in contrast to higher symmetry molecules for which band types are readily deduced by simple symmetry rules. Herein, we present a method to simulate the convoluted rotational contours in the gas-phase spectra of C 1 molecules by combining existing ab initio calculations with Colin Western’s PGOPHER rotational contour program. Specifically, ab initio calculations in the NWC HEM suite of programs were employed to predict the components of the dipole moment derivatives along the principal axes of the moments of inertia. This information was then input into PGOPHER to model the fundamental band contours as a sum of a-, b-, and c-type transitions. This method was applied to simulate the rotational contour spectra of a series of representative C 1 molecules which were then compared against both ab initio stick spectra and experimentally measured broadband IR spectra from the Pacific Northwest National Laboratory infrared gas-phase database. In addition to providing further insight beyond what is revealed in a typical stick spectrum, the simulated contours showed good agreement with the measured spectra.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗