Representation Of Structure Research Articles

Disentangled Representation Learning.

Disentangled Representation Learning (DRL) aims to learn a model capable of identifying and disentangling the underlying factors hidden in the observable data in representation form. The process of separating underlying factors of variation into variables with semantic meaning benefits in learning explainable representations of data, which imitates the meaningful understanding process of humans when observing an object or relation. As a general learning strategy, DRL has demonstrated its power in improving the model explainability, controlability, robustness, as well as generalization capacity in a wide range of scenarios such as computer vision, natural language processing, and data mining. In this article, we comprehensively investigate DRL from various aspects including motivations, definitions, methodologies, evaluations, applications, and model designs. We first present two well-recognized definitions, i.e., Intuitive Definition and Group Theory Definition for disentangled representation learning. We further categorize the methodologies for DRL into four groups from the following perspectives, the model type, representation structure, supervision signal, and independence assumption. We also analyze principles to design different DRL models that may benefit different tasks in practical applications. Finally, we point out challenges in DRL as well as potential research directions deserving future investigations. We believe this work may provide insights for promoting the DRL research in the community.

Process-Based Evaluation of Intraseasonal Oceanic Kelvin Waves in the Pacific Ocean in CMIP6 Models

Abstract Oceanic intraseasonal Kelvin waves (KWs) help modulate upper-ocean thermal characteristics, providing feedbacks to important coupled air–sea phenomena in the tropics. The recent availability of daily thermocline depth fields from several phase 6 of the Coupled Model Intercomparison Project (CMIP6) models makes it possible to evaluate the performance of KWs and identify potential sources of bias. Most models fail to simulate a realistic spatial distribution of KW variability. Models simulate a large variability of KWs in the western or eastern Pacific rather than in the central Pacific as observed. The modeled KWs propagate slowly (about 1.5 m s−1) compared to observations (about 2.5 m s−1). This slow propagation is also identified in wavenumber–frequency spectra for KWs and meridional KW structures, which is more consistent with a second baroclinic mode structure in models compared to the first baroclinic mode structure in observations. An analysis of the relative contributions of the vertical wavenumber and background ocean stability to KW phase speeds indicates that the high vertical wavenumber bias in models contributes most to the slow propagation, in which the higher-than-observed vertical wavenumbers imply the biased incorporation of higher baroclinic modes in the model KW structure. This finding is further supported by the results of vertical mode decomposition that incorporates background density profiles. These results indicate that a realistic representation of the KW vertical structure is essential to produce realistic KW propagations in models. Significance Statement Oceanic intraseasonal Kelvin waves (KWs) play a significant role in regulating the heat content and temperature of the ocean, which provides feedbacks to coupled ocean–atmosphere phenomena in the tropics such as El Niño–Southern Oscillation (ENSO). Consequently, identifying the biases in KWs and the potential sources of those biases in state-of-the-art models is essential to improve simulations of ENSO and its diversity and advance forecasts of weather extremes around the globe induced by ENSO. We find that KWs in the models propagate slower than observations, mainly due to biases in their vertical structure, with secondary effects due to biases in model ocean stability. These issues may be potential sources for current imperfect ENSO model simulations and predictions.

RoadSitu: Leveraging Road Video Frame Extraction and Three-Stage Transformers for Situation Recognition

Situation recognition is an crucial problem in scene understanding, activity understanding, and action reasoning as it provides a structured representation of the main activity depicted in the image.Semantic role labeling is crucial to situation recognition, which is challenging because a single action can have multiple meanings and purposes depending on its context. Understanding images beyond the highlighted actions requires inferences about the context of the scene, the objects, and their role in the captured event. Recently, situation recognition (SR) has been introduced, which jointly derives a collection of the action (activity), meaning-role, and noun (entities) pairs in the form of moving images. To label these frames as action frames, we must assign nouns (entities) to the role based on the content of the observed image. One of the main challenges is managing the complex dependencies between the assigned roles (nouns) and the predicted action, as the correct role assignment often depends on the accuracy of the action prediction. We introduce, RoadSitu, a road situation recognition that involves generating a structured summary of what is happening in a road scenario using an action and the semantic roles played by agents from a video frame. The action can describe a diverse set of situations, and the same agent can play various roles depending on the situation depicted in the video frame. Therefore, a situation recognition model needs to understand the context of each video frame and the visual-linguistic meaning of the semantic roles of that particular frame. One of the main challenges in this work is the complex task of annotating video frames with semantic roles and handling the structured dependencies between the assigned roles (nouns) and the predicted action (activity). Additionally, the sparsity of meaningful semantic information within road scenarios poses further difficulties. To overcome these challenges, we introduce a novel approach where action recognition and noun estimation work together interactively to form structured summaries of each situation. In experiments using a road video dataset obtained from a South Korean company, RoadSitu achieved significant improvements across various performance metrics, with a Top-1 verb accuracy of 43.46%, Top-5 verb accuracy of 72.48%, and value accuracy of 34.21%, outperforming baseline models such as GSRTR and JSL by 2.4% and 3.86% in Top-1 verb accuracy, respectively. These results demonstrate the effectiveness of our model in handling complex road scenarios.

LIA: Latent Image Animator.

Previous animation techniques mainly focus on leveraging explicit structure representations (e.g., meshes or keypoints) for transferring motion from driving videos to source images. However, such methods are challenged with large appearance variations between source and driving data, as well as require complex additional modules to respectively model appearance and motion. Towards addressing these issues, we introduce the Latent Image Animator (LIA), streamlined to animate high-resolution images. LIA is designed as a simple autoencoder that does not rely on explicit representations. Motion transfer in the pixel space is modeled as linear navigation of motion codes in the latent space. Specifically such navigation is represented as an orthogonal motion dictionary learned in a self-supervised manner based on proposed Linear Motion Decomposition (LMD). Extensive experimental results demonstrate that LIA outperforms state-of-the-art on VoxCeleb, TaichiHD, and TED-talk datasets with respect to video quality and spatio-temporal consistency. In addition LIA is well equipped for zero-shot high-resolution image animation. Code, models, and demo video are available at https://github.com/wyhsirius/LIA.

The difference in image quality assessment between radiographers and radiologists and its relationship with diagnostic accuracy.

Radiographers must ensure that the radiographs they produce are of sufficient quality before reporting by radiologists. However, does radiographers' perception of image quality relates to the diagnostic accuracy of radiologists? 38 unique postero-anterior (PA) chest radiographs containing either no pathology, well defined or diffuse lesions were presented to four radiologists and 24 radiographers. Image quality was rated using RadLex quality descriptors. Additionally, radiologists rated their confidence of a radiograph being were normal or abnormal on a 5-point scale. Radiographers rated visibility of anatomical structures, satisfaction with positioning, collimation and adequacy of detector dose. Modelling correlation explored relationships between radiologists' detection of pathology and radiographers' ratings. Radiographers rated image quality of 56% of radiographs as non-diagnostic or limited, compared to 30.7% by radiologists. Modelling revealed that chances of a radiograph being classified in lower quality categories by radiographers increased inversely proportionally with perceived dose, and suddenly increased when positioning was considered unacceptable. No correlation (p=0.10) was found between quality descriptors by radiographers and diagnostic accuracy of radiologists. Findings suggest that perceived "better" image quality, as defined by radiographers, may not necessarily correlate with diagnostic accuracy. This implies that radiographs repeated in clinical practice might not enhance diagnostic value. These findings raise the question which elements in the radiography or actual content (such as representation of anatomical structures) are key elements used to diagnose, which may have implications for image quality studies. Exploring the evolution of decision-making in radiographers trained to report could provide further valuable insights.

Dynamic graph structure and spatio-temporal representations in wind power forecasting

Wind Power Forecasting (WPF) has gained considerable focus as a crucial aspect of the successful integration and operation of wind power. However, due to the stochastic and unstable nature of wind, it poses a real challenge to effectively analyze the correlations among multiple time series data for accurate prediction. In our study, an end-to-end framework called Dynamic Graph structure and Spatio-Temporal representation learning (DSTG) framework is proposed to achieve stable power forecasting by constructing graph data to capture the critical features in the data. Specifically, a graph structure learning (GSL) module is introduced to dynamically construct task-related correlation matrices via backpropagation to mitigate the inherent inconsistency and randomness of wind power data. Additionally, a dual-scale temporal graph learning (DTG) module is further proposed to explore the implicit spatio-temporal features at a fine-grained level using different skip connections from the constructed graph data. Finally, comprehensive experiments are performed on the collected Xuji Group Wind Power (XGWP) dataset, and the results show that DSTG outperforms the state-of-the-art spatio-temporal methods by 10.12% on the average of root mean square error and mean absolute error, demonstrating the effectiveness of DSTG. In conclusion, our model provides a promising approach.

STRUCTURE-ACTIVITY RELATIONSHIP OF PSYCHOACTIVE CANNABINOIDS: A CHEMOMETRIC EXPERIMENT FOR UNDERGRADUATE AND POSTGRADUATE STUDENTS IN CHEMICAL EDUCATION

Undergraduate and postgraduate Chemistry Education students had to develop the experiment in four (4) stages. Twenty-two (22) cannabinoids (Δ9-THC and derivatives) with different degrees of psychoactivity, previously scrutinized by our research group with the approaches: molecular electrostatic potential (MEP) and interaction with the CB1 receptor (CARDOSO-FILHO et al., 2004), are investigated through Chemometrics. Exploratory Analysis: Principal Component Analysis (PCA) and Hierarchical Cluster Analysis (HCA). E Classification Methods: K-Nearest Neighbor (KNN) method, Soft Independent Modeling Class Analogy (SIMCA) method, and Stepwise Discriminant Analysis (SDA) were employed to reduce the dimensionality of the data matrix and investigate which descriptors are responsible for the classification between the most active cannabinoids (mac), pKi ≥ 6.72, and the least active cannabinoids (lac), pKi < 6.72, according to the hypothesis previously reported (Cardoso-Filho et al., 2024). The investigation with PCA, HCA, KNN, SIMCA, and SDA showed that the descriptors LUMO+1 energy, Geary autocorrelation of lag 6 weighted by van der Waals volumes (GATS6V), hydration energy (EH), atomic charge on the atom of C4 (qC4), and molecular representation of structure based on the electron diffraction, code of signal 3, unweighted (MOR03u) are responsible for separating cannabinoids according to their degrees of psychoactivity. The insights accumulated and the models built in the research – PCA chemometric model, HCA chemometric model, KNN chemometric model, SIMCA chemometric model, and SDA chemometric model – will be able to support the design of new potentially psychoactive Δ9-THC derivatives.

Semantic and distributional patterns of Spanish negation with nouns and adjectives: A Lexical-Realizational Functional Grammar approach

This study examines the semantic and distributional characteristics of the Spanish negation no when it occurs before nouns and adjectives, specifically, whether these instances diverge from sentential negation, but also from negation via prefixes such as des- ‘dis-’ or in- ‘in-, un-’. Using data from the Spanish Web 2011 corpus, it is shown that the use of no before adjectives aligns with other forms of constituent negation, often resulting in a contradictory interpretation, that is similar to that of sentential negation. However, the interpretation of no preceding nouns exhibits a broader range of interpretations, depending on whether the noun refers to an eventuality, quality, or entity. In such instances, no shows parallels to negative prefixes, frequently indicating privation rather than contradiction. Consequently, no can be analyzed as a building block of phrases, expressing syntactic negation, behaving similarly to other adverbs, as well as a building block of words, expressing lexical negation, resembling Spanish prefix-like elements like cuasi ‘almost’ or ex ‘ex-’. The analysis is couched within the framework of Lexical-Realizational Functional Grammar (LRFG), integrating aspects of Lexical Functional Grammar and Distributed Morphology. This approach does not treat morphology and syntax as separate modules, thus allowing us to account for the variable distribution and interpretation of no by means of a single representation in vocabulary structure, drawing on constituent structure representations which are needed on independent grounds for adverbs and for prefix-like elements.

Characterization of entropy measures with connection number based indices of boric acid hydrogen-bonded 2D lattice sheets.

The cut method is a computational approach utilized to predict the fundamental activities of physicochemical properties of chemical networks, also called topological indices. The connection number is a new idea, that gives interesting and good results of the topological indices (TIs) and entropy measures (EMs) for structural representation of chemical compounds and networks. The physical density of chemical networks is characterized by these indices. In this paper, we determined the computational results for indices based on connection numbers for a two-dimensional lattice sheet of hydrogen-bonded boric acid. Boric acid, an inorganic compound, is not very harmful when applied to the skin or consumed. Finally, graphical and numerical comparisons of topological numbers including the number of borate hydrogen-bonded double lattice forms are also included in this study.

DeepMEns: an ensemble model for predicting sgRNA on-target activity based on multiple features.

The CRISPR/Cas9 system developed from Streptococcus pyogenes (SpCas9) has high potential in gene editing. However, its successful application is hindered by the considerable variability in target efficiencies across different single guide RNAs (sgRNAs). Although several deep learning models have been created to predict sgRNA on-target activity, the intrinsic mechanisms of these models are difficult to explain, and there is still scope for improvement in prediction performance. To overcome these issues, we propose an ensemble interpretable model termed DeepMEns based on deep learning to predict sgRNA on-target activity. By using five different training and validation datasets, we constructed five sub-regressors, each comprising three parts. The first part uses one-hot encoding, wherein 0-1 representation of the secondary structure is used as the input to the convolutional neural network (CNN) with Transformer encoder. The second part uses the DNA shape feature matrix as the input to the CNN with Transformer encoder. The third part uses positional encoding feature matrices as the proposed input into a long short-term memory network with an attention mechanism. These three parts are concatenated through the flattened layer, and the final prediction result is the average of the five sub-regressors. Extensive benchmarking experiments indicated that DeepMEns achieved the highest Spearman correlation coefficient for 6 of 10 independent test datasets as compared to previous predictors, this finding confirmed that DeepMEns can accomplish state-of-the-art performance. Moreover, the ablation analysis also indicated that the ensemble strategy may improve the performance of the prediction model.

Simulating Ips typographus L. outbreak dynamics and their influence on carbon balance estimates with ORCHIDEE r8627

Abstract. New (a)biotic conditions resulting from climate change are expected to change disturbance dynamics, such as windthrow, forest fires, droughts, and insect outbreaks, and their interactions. These unprecedented natural disturbance dynamics might alter the capability of forest ecosystems to buffer atmospheric CO2 increases, potentially leading forests to transform from sinks into sources of CO2. This study aims to enhance the ORCHIDEE land surface model to study the impacts of climate change on the dynamics of the bark beetle, Ips typographus, and subsequent effects on forest functioning. The Ips typographus outbreak model is inspired by previous work from Temperli et al. (2013) for the LandClim landscape model. The new implementation of this model in ORCHIDEE r8627 accounts for key differences between ORCHIDEE and LandClim: (1) the coarser spatial resolution of ORCHIDEE; (2) the higher temporal resolution of ORCHIDEE; and (3) the pre-existing process representation of windthrow, drought, and forest structure in ORCHIDEE. Simulation experiments demonstrated the capability of ORCHIDEE to simulate a variety of post-disturbance forest dynamics observed in empirical studies. Through an array of simulation experiments across various climatic conditions and windthrow intensities, the model was tested for its sensitivity to climate, initial disturbance, and selected parameter values. The results of these tests indicated that with a single set of parameters, ORCHIDEE outputs spanned the range of observed dynamics. Additional tests highlighted the substantial impact of incorporating Ips typographus outbreaks on carbon dynamics. Notably, the study revealed that modeling abrupt mortality events as opposed to a continuous mortality framework provides new insights into the short-term carbon sequestration potential of forests under disturbance regimes by showing that the continuous mortality framework tends to overestimate the carbon sink capacity of forests in the 20- to 50-year range in ecosystems under high disturbance pressure compared to scenarios with abrupt mortality events. This model enhancement underscores the critical need to include disturbance dynamics in land surface models to refine predictions of forest carbon dynamics in a changing climate.

Accuracy Assessment of an Image-Based Cloud-Based Indoor Mapping Platform as an Alternative for Desktop Applications

Abstract. The generation of 3D spatial datasets, especially in indoor space, has continually been a challenge in developing spatial applications that are aimed for the realistic representation of the real world. Despite their potential, the creation of 3D spatial data remains a challenge, particularly in terms of cost and complexity associated with traditional methods. While this method proves to be a reliable and accurate method of creating such data, this approach entails challenges in economic, computational, and human resource aspects. Alternatively, omnidirectional images have emerged as cost-effective alternatives for representing 3D spaces, offering a comprehensive field of view and facilitating the generation of structure and appearance representations through mesh-based 3D maps. This study evaluates the geometric accuracy of a cloud-based image-based mapping platform for generating 3D representations of indoor spaces, aiming to assess its suitability as an alternative to desktop-based platforms. We obtained 30 sample lines and compare respective measurements from each of the generated meshes to a ground truth value. Results show that while there are no significant differences in the mean errors, the mesh generated from a cloud-based platform produced minimal errors. This demonstrates the potential of this platform for generating 3D representations of indoor space, acceptable in both geometric accuracy and visualization capabilities.

A case study for automated attribute extraction from legal documents using large language models

AbstractThe escalating number of pending cases is a growing concern worldwide. Recent advancements in digitization have opened up possibilities for leveraging artificial intelligence (AI) tools in the processing of legal documents. Adopting a structured representation for legal documents, as opposed to a mere bag-of-words flat text representation, can significantly enhance processing capabilities. With the aim of achieving this objective, we put forward a set of diverse attributes for criminal case proceedings. To enhance the effectiveness of automatically extracting these attributes from legal documents within a sequence labeling framework, we propose the utilization of a few-shot learning approach based on Large Language Models (LLMs). Moreover, we demonstrate the efficacy of the extracted attributes in downstream tasks, such as legal judgment prediction and legal statute prediction.

Multiscale attention-over-attention network for retinal disease recognition in OCT radiology images

Retinal disease recognition using Optical Coherence Tomography (OCT) images plays a pivotal role in the early diagnosis and treatment of conditions. However, the previous attempts relied on extracting single-scale features often refined by stacked layered attentions. This paper presents a novel deep learning-based Multiscale Feature Enhancement via a Dual Attention Network specifically designed for retinal disease recognition in OCT images. Our approach leverages the EfficientNetB7 backbone to extract multiscale features from OCT images, ensuring a comprehensive representation of global and local retinal structures. To further refine feature extraction, we propose a Pyramidal Attention mechanism that integrates Multi-Head Self-Attention (MHSA) with Dense Atrous Spatial Pyramid Pooling (DASPP), effectively capturing long-range dependencies and contextual information at multiple scales. Additionally, Efficient Channel Attention (ECA) and Spatial Refinement modules are introduced to enhance channel-wise and spatial feature representations, enabling precise localization of retinal abnormalities. A comprehensive ablation study confirms the progressive impact of integrated blocks and attention mechanisms that enhance overall performance. Our findings underscore the potential of advanced attention mechanisms and multiscale processing, highlighting the effectiveness of the network. Extensive experiments on two benchmark datasets demonstrate the superiority of the proposed network over existing state-of-the-art methods.

IFGAN: Pre- to Post-Contrast Medical Image Synthesis Based on Interactive Frequency GAN

Medical images provide a visual representation of the internal structure of the human body. Injecting a contrast agent can increase the contrast of diseased tissues and assist in the accurate identification and assessment of conditions. Considering the adverse reactions and side effects caused by contrast agents, previous methods synthesized post-contrast images with pre-contrast images to bypass the administration process. However, existing methods pay inadequate attention to reasonable mapping of the lesion area and ignore gaps between post-contrast and real images in the frequency domain. Thus, in this paper, we propose an interactive frequency generative adversarial network (IFGAN) to solve the above problems and synthesize post-contrast images from pre-contrast images. We first designed an enhanced interaction module that is embedded in the generator to focus on the contrast enhancement region. Within it, target and reconstruction branch features interact to control the local contrast enhancement region feature and maintain the anatomical structure. We propose focal frequency loss to ensure the consistency of post-contrast and real images in the frequency domain. The experimental results demonstrated that IFGAN outperforms other sophisticated approaches in terms of preserving the accurate contrast enhancement of lesion regions and anatomical structures. Specifically, our method produces substantial improvements of 7.9% in structural similarity (SSIM), 36.3% in the peak signal-to-noise ratio (PSNR), and 8.5% in multiscale structural similarity (MSIM) compared with recent state-of-the-art methods.

Perception of the Urban Space of the Waorani Nationality in Shell Township, Pastaza, Ecuador

The township of Shell in the Province of Pastaza, Ecuador was initially an exploration base and later, a missionary base for evangelization of the peoples and nationalities of the Amazon. It plays an important role in the Waorani nationality and its social, cultural and economic dynamics, especially in the Nuevo Amanecer neighborhood, which is home to approximately 43 Waorani families whose epicenter has become a strategic space for their territorial dynamics. The research, through ethnographic methodologies and field work, has allowed to establish the perception from the Waorani nationality to the urban space. During the development of this study two sources of information were applied: methodological bibliography and thematic content reference. The study of 120 members of 43 Waorani families of the Ecuadorian Amazon and a process of direct observation, as a participant observer during the years 2019-2020-2021-2022, resulted in a conceptual model Bailey (1996), where the Waorani use preferences of displacement, economic attributes of the area, location of residence, representation of the spatial-urban structure, space-action, time of stay, activity space and objective spatial structure of the urban environment; and a transect as a very effective method to recognize population movements, sectoral differences, topographic slopes and social dynamics presented in the territory from a plan and elevation view. Human beings adapt easily to new spaces, in spite of cultural and social differences, their assimilation processes will depend on their capacity to approach and accept the environment. The Nuevo Amanecer settlement has motivated families to move to peri-urban or urban spaces in intermediary cities, which has provoked a forced adaptation and assimilation due to the same processes involved in rural-urban relations. Received: 30 May 2024 / Accepted: 21 October 2024 / Published: 05 November 2024

Bi-directional information interaction for multi-modal 3D object detection in real-world traffic scenes

Multimodal 3D object detection methods are poorly adapted to real-world traffic scenes due to sparse distribution of point clouds and misalignment multimodal data during actual collection. Among the existing methods, they focus on high-quality open-source datasets, with performance relying on the accurate structural representation of point clouds and the precise mapping relationship between point clouds and images. To solve the above challenges, this paper proposes a multimodal feature-level fusion method based on the bi-directional interaction between image and point cloud. To overcome the sparsity issue in asynchronous multi-modal data, a point cloud densification scheme based on visual guidance and point cloud density guidance is proposed. This scheme can generate object-level virtual point clouds even when the point cloud and image are misaligned. To deal with the unalignment issue between point cloud and image, a bi-directional interaction module based on image-guided interaction with key points of point clouds and point cloud-guided interaction with image context information is proposed. It achieves effective feature fusion even when the point cloud and image are misaligned. The experiments on the VANJEE and KITTI datasets demonstrated the effectiveness of the proposed method, with average precision improvements of 6.20% and 1.54% compared to the baseline.

Probing visual literacy skills reveals unexpected student conceptions of chromosomes.

Molecular biology can be challenging for undergraduate students because it requires visual literacy skills to interpret abstract representations of submicroscopic concepts, structures and processes. The Conceptual-Reasoning-Mode framework suggests that visual literacy relies on applying conceptual knowledge to appropriately reason with the different ways of representing concepts in molecular biology. We used this framework to specifically explore visual literacy related to chromosomes. We conducted 35 semi-structured interviews with students who had taken at least a year of college-level biology courses, and we asked them to sketch chromosomes, interpret an abstract representation of chromosomes, and use the abstract representation to answer a multiple-choice question about meiosis. While many participants used the correct vocabulary to describe chromosome structure and function, probing their visual literacy skills revealed gaps in their understanding. Notably, 97% of participants (34 of 35) held conceptual errors related to chromosome structure and function, which were often only revealed in their sketches or explanations of their sketches. Our findings highlight the importance of scaffolding visual literacy skills into instruction by teaching with a variety of visual models and engaging students in using and interpreting the conventions of abstract representations of chromosomes.

Dynamic analysis of a pedestrian bridge using monitoring and computational techniques

Soil-structure interaction (SSI) represents an important parameter when analyzing the dynamic behavior of bridges, considering the combined effects of soil, foundation, and structure to ensure an adequate structural representation. This paper applied a computational modeling strategy associated with soil-structure interaction methodology to represent the dynamic behavior of a timber-concrete pedestrian bridge of the Polytechnic School of the University of São Paulo. The computational models were validated and calibrated using monitoring results from an extensive structural health monitoring campaign. The bridge monitoring program was characterized using embedded sensors and remote monitoring technology: ground-based radar interferometry and video motion amplification. The instrumentation results were compared to establish the capabilities and advantages of this novel remote monitoring technique. The experimental results obtained from free and forced vibrations in the asset served as a basis for validating the developed computational model and allowed the evaluation of the asset behavior in service for different loading conditions. This paper emphasizes the importance of considering an SSI approach to represent the bridge dynamic behavior more accurately, enabling the use of more reliable numerical models to assess the bridge structural responses throughout its life-cycle.

Unsupervised representation learning of Kohn–Sham states and consequences for downstream predictions of many-body effects

Representation learning for the electronic structure problem is a major challenge of machine learning in computational condensed matter and materials physics. Within quantum mechanical first principles approaches, density functional theory (DFT) is the preeminent tool for understanding electronic structure, and the high-dimensional DFT wavefunctions serve as building blocks for downstream calculations of correlated many-body excitations and related physical observables. Here, we use variational autoencoders (VAE) for the unsupervised learning of DFT wavefunctions and show that these wavefunctions lie in a low-dimensional manifold within latent space. Our model autonomously determines the optimal representation of the electronic structure, avoiding limitations due to manual feature engineering. To demonstrate the utility of the latent space representation of the DFT wavefunction, we use it for the supervised training of neural networks (NN) for downstream prediction of quasiparticle bandstructures within the GW formalism. The GW prediction achieves a low error of 0.11 eV for a combined test set of two-dimensional metals and semiconductors, suggesting that the latent space representation captures key physical information from the original data. Finally, we explore the generative ability and interpretability of the VAE representation.