SEARCH · Search NASA

Results for “feature set”

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.

SourceResearch connectionFrom yearThrough year

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

At least 397 records · Page 22

All sources NASA NTRS DOE OSTI

DOE OSTI2024

MolSym : A Python package for handling symmetry in molecular quantum chemistry

A consideration of the point group symmetry of molecules is often advantageous from a computational efficiency standpoint and sometimes necessary for the correct treatment of chemical physics problems. Many modern electronic structure software packages include a treatment of symmetry, but these are sometimes incomplete or unusable outside of that program’s environment. Therefore, we have developed the MolSym package for handling molecular symmetry and its associated functionalities to provide a platform for including symmetry in the implementation and development of other methods. Features include point group detection, molecule symmetrization, arbitrary generation of symmetry element sets and character tables, and symmetry adapted linear combinations of real spherical harmonic basis functions, Cartesian displacement coordinates, and internal coordinates. We present some of the advantages of using molecular symmetry as achieved by MolSym, particularly with respect to Hartree–Fock theory, and the reduction of finite difference displacements in gradient/Hessian computations. Furthermore, this package is designed to be easily integrated into other software development efforts and may be extended to further symmetry applications.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

DOE OSTI2020

Improving galaxy clustering measurements with deep learning: analysis of the DECaLS DR7 data

ABSTRACT Robust measurements of cosmological parameters from galaxy surveys rely on our understanding of systematic effects that impact the observed galaxy density field. In this paper, we present, validate, and implement the idea of adopting the systematics mitigation method of artificial neural networks for modelling the relationship between the target galaxy density field and various observational realities including but not limited to Galactic extinction, seeing, and stellar density. Our method by construction allows a wide class of models and alleviates overtraining by performing k-fold cross-validation and dimensionality reduction via backward feature elimination. By permuting the choice of the training, validation, and test sets, we construct a selection mask for the entire footprint. We apply our method on the extended Baryon Oscillation Spectroscopic Survey (eBOSS) Emission Line Galaxies (ELGs) selection from the Dark Energy Camera Legacy Survey (DECaLS) Data Release 7 and show that the spurious large-scale contamination due to imaging systematics can be significantly reduced by up-weighting the observed galaxy density using the selection mask from the neural network and that our method is more effective than the conventional linear and quadratic polynomial functions. We perform extensive analyses on simulated mock data sets with and without systematic effects. Our analyses indicate that our methodology is more robust to overfitting compared to the conventional methods. This method can be utilized in the catalogue generation of future spectroscopic galaxy surveys such as eBOSS and Dark Energy Spectroscopic Instrument (DESI) to better mitigate observational systematics.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2026

Characterizing skyrmion flow phases with principal component analysis

Principal component analysis (PCA) is a powerful method that can identify patterns in large, complex data sets by constructing low-dimensional order parameters from higher-dimensional feature vectors. There are increasing efforts to use space-and-time-dependent PCA to detect transitions in nonequilibrium systems that are difficult to characterize with equilibrium methods. Here, we demonstrate that feature vectors incorporating the position and velocity information of driven skyrmions moving through random disorder permit PCA to resolve different types of disordered skyrmion motion as a function of driving force and the ratio of the Magnus force to the dissipation. Since the Magnus force creates gyroscopic motion and a finite Hall angle, skyrmions can exhibit a greater range of flow phases than what is observed in overdamped driven systems with quenched disorder. We show that in addition to identifying previously known skyrmion flow phases, PCA detects several additional phases, including different types of channel flow, moving fluids, and partially ordered states. Guided by the PCA analysis, we further characterize the disordered flow phases to elucidate the different microscopic dynamics and show that the changes in the PCA-derived order parameters can be connected to features in bulk transport measures, including the transverse and longitudinal velocity-force curves, differential conductivity, topological defect density, and changes in the skyrmion Hall angle as a function of drive. We discuss how asymmetric feature vectors can be used to improve the resolution of the PCA analysis, and how this technique can be extended to find disordered phases in other nonequilibrium systems with time-dependent dynamics.

36 MATERIALS SCIENCE↗

DOE OSTI2025

On the minimum number of radiation field parameters to specify gas cooling and heating functions

Fast and accurate approximations of gas cooling and heating functions are needed for hydrodynamic galaxy simulations. We use machine learning to analyze atomic gas cooling and heating functions in the presence of a generalized incident local radiation field computed by Cloudy. We characterize the radiation field through binned radiation field intensities instead of the photoionization rates used in our previous work. We find a set of 6 energy bins whose intensities exhibit relatively low correlation. We use these bins as features to train machine learning models to predict Cloudy cooling and heating functions at fixed metallicity. We compare the relative SHapley Additive exPlanation (SHAP) value importance of the features. From the SHAP analysis, we identify a feature subset of 3 energy bins (0.5-1, 1-4, and 13-16Ry) with the largest importance and train additional models on this subset. We compare the mean squared errors and distribution of errors on both the entire training data table and a randomly selected 20% test set withheld from model training. The machine learning models trained with 3 and 6 bins, as well as 3 and 4 photoionization rates, have comparable accuracy everywhere, with errors ≳10 times smaller than for the interpolation table of Gnedin and Hollon (2012). We conclude that 3 energy bins (or 3 analogous photoionization rates: molecular hydrogen photodissociation, neutral hydrogen HI, and fully ionized carbon CVI) are sufficient to characterize the dependence of the gas cooling and heating functions on our assumed incident radiation field model.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2023

Informed Feature Selection for Data Clustering of CSP Plant Production

To make concentrating solar power (CSP) more cost competitive, rigourous optimizations must be run to improve plant design and operations. However, these optimizaitons rely on time consuming annual simulations that solve an electricity dispatch scheduling problem to maximize plant revenue. To reduce the runtime of annual dispatch simulations of CSP plants, a data clustering approach is utilized. This approach assumes that like days of revenue and electricity generation can be identified using weather and price data. Although weather and price are important factors for electricity production, this work investigates how thermal energy storage (TES) inventory at the beginning of a day, denoted as Si, can be used as a supplemental feature to group like days. A framework for creating and training a deep neural network to predict Si is proposed. This model is validated and assessed using eleven sets of testing data that were not used during training. Then, the data clustering approach is performed three seperate times with features of weather and price along with either Si from the neural network, Si from the full annual simulation, or no Si. Ultimately, the results suggest that using Si as an additional clustering feature improves the data clustering simulation accuracy by 1.4%.

Tuman, Matthew J. (ORCID:000900038772051X)↗

DOE OSTI2021

System and method for structural characterization of materials by supervised machine learning-based analysis of their spectra

A method of supervised machine learning-based spectrum analysis information, using a neural network trained with spectrum information, to identify a specified feature of a given material, a system for supervised machine learning-based spectrum analysis, and a method of training a neural network to analyze spectrum data. The method of supervised machine learning-base spectrum analysis comprises inputting into the neural network spectrum data obtained from a sample of the given material; and the neural network processing the spectrum data, in accordance with the training of the neural network, and outputting one or more values for the specified feature of the sample of the material. In an embodiment, the training set of data includes x-ray absorption spectroscopy data for the given material. In an embodiment, the training set of data includes electron energy loss spectra (EELS) data.

Frenkel, Anatoly↗

DOE OSTI2020

Development of Gamma Background Radiation Digital Twin with Machine Learning Algorithms: Application of Unsupervised Machine Learning to Detection of Anomalies and Nuisances in Gamma Background Radiation Environmental Screening Data

Environmental screening of gamma radiation consists of detecting weak nuisance and anomaly signal in the presence of strong and highly varying background. In a typical scenario, a mobile detector-spectrometer continuously measures gamma radiation spectra in short, e.g., one-second, signal acquisition intervals. The measurement data is a 2D matrix, where one dimension is gamma ray energy, and the other dimension is the number of measurements or total time. In principle, gamma radiation sources can be detected and identified from the measured data by their unique spectral lines. Detecting sources from data measured in a search scenario is difficult due to the highly varying background because of naturally occurring radioactive material (NORM), and low signal-to-noise ratio (S/N) of spectral signal measured during one-second acquisition intervals. The objective of this work is to explore unsupervised machine learning (ML) algorithms for development of a digital twin of gamma radiation background, and for detection and identification of weak nuisances and anomalies events in the presence of highly fluctuating background. In one segment of work, we developed a gamma background estimation model using a Longshort term memory (LSTM) network for one-step CPS time series prediction. The LSTM model was validated with two data sets of measurements from two independent NaI detectors positioned on a mobile platform. The data sets contained background radiation only and no orphan isotope sources. The LSTM model was constructed and tested using data from one of the detectors. Performance of the LSTM model was validate through one-step prediction of CPS time series of another NaI detector without re-training. This approach allows to create a digital twin for nuclear background estimation. Using LSTM, it could be possible to detect a source through subtraction of the estimated counts from the measured background. In another segment of work, we investigated detection of gamma emitting sources in the presence of complex background using unsupervised machine learning. Spectral lines of isotopes are difficult to observe in one-second measurements. Averaging over the entire measurement campaign data set reveals spectral lines of most common background isotopes. Spectral lines of orphan sources, which might appear only in a few measurements during the campaign, will be washed out if averaging is performed over the entire measurement data set. The approach we have explored consists of extracting one-second measurements containing weak spectral features through data clustering. Averaging one-second spectra in a cluster should reveal the presence of anomaly sources. We created two ML models using K-means clustering and Neural Network Self-organizing Map (SOM). Performance of these ML models was benchmarked using search data. One data set contained 137 Cs source, and another dataset contained 131 I source.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2021

Detection of Anomalies in Gamma Background Radiation Data with K-Means and Self-Organizing Map Clustering Algorithms (Consortium on Nuclear Security Technologies (CONNECT) Q1 Report)

61 RADIATION PROTECTION AND DOSIMETRY↗

DOE OSTI2025

On the minimum number of radiation field parameters to specify gas cooling and heating functions

Fast and accurate approximations of gas cooling and heating functions are needed for hydrodynamic galaxy simulations. We use machine learning to analyze atomic gas cooling and heating functions computed by Cloudy in the presence of a generalized incident local radiation field. We characterize the radiation field through binned radiation field intensities instead of the photoionization rates used in our previous work. We find a set of 6 energy bins whose intensities exhibit relatively low correlation. We use these bins as features to train machine learning models to predict Cloudy cooling and heating functions at fixed metallicity. We compare the relative SHapley Additive exPlanation (SHAP) value importance of the features. From the SHAP analysis, we identify a feature subset of 3 energy bins ($0.5-1, 1-4$, and $13-16 \, \mathrm{Ry}$) with the largest importance and train additional models on this subset. We compare the mean squared errors and distribution of errors on both the entire training data table and a randomly selected 20% test set withheld from model training. The machine learning models trained with 3 and 6 bins, as well as 3 and 4 photoionization rates, have comparable accuracy everywhere, with errors $\gtrsim 10$ times smaller than for the interpolation table of Gnedin and Hollon (2012). We conclude that 3 energy bins (or 3 analogous photoionization rates: molecular hydrogen photodissociation, neutral hydrogen HI, and fully ionized carbon CVI) are sufficient to characterize the dependence of the gas cooling and heating functions on our assumed incident radiation field model.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2020

Computationally Aided Design of a Small-Scale Waste-Glass Melter - 20452

To provide mission support to the WTP for the vitrification of legacy nuclear tank waste at the Hanford site, a reduced-scale vitrification pilot system is being designed to process simulated and actual radioactive tank wastes. The waste stored in underground tanks will be separated into HLW and LAW fractions, where the majority is LAW by mass (approximately 90%) and HLW by activity (greater than 95%). The first tank waste to be processed at the WTP will be LAW. Because Hanford tank waste and the resultant melter feed compositions are known to vary widely, pilot-scale operations are essential to identify potential problems, provide needed data, confirm assumptions, and determine impacts to full-scale melters and off-gas systems from these different feeds. In addition, the reduced-scale melter system must be capable of quickly providing results to operations without the typically high cost of full-scale radioactive operations and to provide an engineering platform in close proximity to local operations staff. The objectives of this computational modeling effort are to produce process conditions like those encountered in the WTP LAW melters and to minimize the volume of radioactive waste necessary for the evaluations. To this end, melter design features are being explored to increase production without significantly increasing surface area or melter-glass volumes or compromising data quality. In support of these objectives, a set of computational fluid dynamic models was developed to provide insight into the flow patterns within these small melters and evaluate design features to increase production without significantly increasing glass inventory. Computed velocities were evaluated at the interface between the cold cap and the molten-glass pool and within the bulk-glass pool. Since melt rate correlates with the velocity underneath the cold cap, this information can be used to select design features that produce the highest gains. (authors)

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

DOE OSTI2023

A Level Set Model for Structured Mesh Representation of Fibrous Preforms used in Composites Manufacturing

Here, a model for implicit representation of fibrous and woven preforms used for composites manufacturing is described. The method is based on a level set function defined on a structured mesh to implicitly capture the complex fiber and weave geometries. Since most software packages used to model woven preforms are based on unstructured mesh, a numerical model for transforming a discretized surface to a structured mesh representation and vice-a-versa is presented. Specific emphasis is on scalable computation of a level set function for large and complex 3D woven structures from a triangulated surface of the smallest unit (yarn). The model is applied to capture a periodic 3D layered 5-harness satin (5HS) weave geometry on a structured mesh. Geometric transformations needed to efficiently compute the level set function for the 3D structure on a high resolution mesh are described. Effects of mesh resolution on the geometric features such as number of resolved fibers, effective fiber size, and the surface area to volume ratio of the iso-surface are analyzed. Using the distance property of the level set function, an approach to evaluate the structure function resulting from transient evolving topology representing densification due to matrix phase depositions is also presented. The approach captures transition of the geometric characteristic from fiber governed at early stages of densification to yarn governed at the terminal stage.

36 MATERIALS SCIENCE↗

DOE OSTI2022

A round-robin approach provides a detailed assessment of biomolecular small-angle scattering data reproducibility and yields consensus curves for benchmarking

Through an expansive international effort that involved data collection on 12 small-angle X-ray scattering (SAXS) and four small-angle neutron scattering (SANS) instruments, 171 SAXS and 76 SANS measurements for five proteins (ribonuclease A, lysozyme, xylanase, urate oxidase and xylose isomerase) were acquired. From these data, the solvent-subtracted protein scattering profiles were shown to be reproducible, with the caveat that an additive constant adjustment was required to account for small errors in solvent subtraction. Further, the major features of the obtained consensus SAXS data over the q measurement range 0–1 Å −1 are consistent with theoretical prediction. The inherently lower statistical precision for SANS limited the reliably measured q -range to <0.5 Å −1 , but within the limits of experimental uncertainties the major features of the consensus SANS data were also consistent with prediction for all five proteins measured in H 2 O and in D 2 O. Thus, a foundation set of consensus SAS profiles has been obtained for benchmarking scattering-profile prediction from atomic coordinates. Additionally, two sets of SAXS data measured at different facilities to q > 2.2 Å −1 showed good mutual agreement, affirming that this region has interpretable features for structural modelling. SAS measurements with inline size-exclusion chromatography (SEC) proved to be generally superior for eliminating sample heterogeneity, but with unavoidable sample dilution during column elution, while batch SAS data collected at higher concentrations and for longer times provided superior statistical precision. Careful merging of data measured using inline SEC and batch modes, or low- and high-concentration data from batch measurements, was successful in eliminating small amounts of aggregate or interparticle interference from the scattering while providing improved statistical precision overall for the benchmarking data set.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2025

MS25: Materials Science-Focused Benchmark Data Set for Machine Learning Interatomic Potentials

Here, we present MS25, a benchmark data set for evaluating machine learning interatomic potentials (MLIPs) across diverse materials-relevant systems including MgO surfaces, liquid water, zeolites, a catalytic Pt surface reaction, high-entropy alloys (HEAs), and disordered Zr-oxides. Five MLIP architectures (MACE, NequIP, Allegro, MTP, and Torch-ANI) are trained and tested, focusing not only on traditional metrics (energies, forces, and stresses) but also explicitly validating derived physical observables such as lattice constants, volumes, and reaction barriers. We find that most models reach comparable accuracy on standard error metrics across the simple systems, although equivariant MLIPs offer 1.5–2× improvements over nonequivariant MLIPs in energy and force error for structurally complex or compositionally disordered environments such as HEAs and Zr–O systems. Our analysis highlights that low errors in energy and force predictions do not guarantee reliable observables, emphasizing the necessity of explicit validation. We demonstrate limitations in cross-framework transferability, as models trained on one zeolite framework (CHA) fail to reliably generalize to predictions of structurally distinct frameworks (e.g., MFI). Size-extensive tests show some dependence on system size for MgO, resulting from forced periodicity. The HEA and Zr–O data sets are identified as challenging tests for future benchmarks and MLIP model architecture developments as they show significant differentiation in error between MLIP architectures and are still relatively difficult at 1000 training images. Moving forward, we recommend that benchmarking efforts shift their focus from marginal accuracy improvements in energy and force errors toward identifying and understanding model failure modes, rigorously assessing transferability, and evaluating how their errors affect observable predictions. For researchers looking to choose an MLIP architecture, we suggest selecting equivariant MLIP architectures if the complexity of the system is a challenge. For simple materials problems, auxiliary features such as integration with molecular dynamics engines, trade-offs between computational data set generation cost vs MLIP inference speed, and framework integration may play a more important decision factor than small differences in error metrics that are unlikely to matter for production-level research.

chemical structure↗

DOE OSTI2023

Empirical radius formulas for canonical neutron stars from bidirectionally selecting features of equations of state in extended Bayesian analyses of observational data

Significant advancement in Bayesian inference of nuclear equation of state (EOS) from gravitational wave and x-ray observations of neutron stars (NSs) has been made by the nuclear astrophysics community especially since GW170817. By extending the traditional Bayesian analysis which normally ends at presenting the marginalized posterior probability distribution functions (PDFs) of individual EOS parameters and their correlations (or sometimes only the Pearson correlation coefficients which are only reliably useful when the variables are linearly correlated while they are actually often not), we search for a data-driven and robust empirical formula for the radius 𝑅 1.4 of canonical NSs in terms of the characteristic EOS parameters (features). We also identify the single most important but currently poorly known EOS parameter for determining the 𝑅 1.4 . Using three regression-model-building methodologies: bidirectional stepwise feature selection, least absolute shrinkage selection operator (LASSO) regression, and neural network regression on a large set of posterior EOSs and the corresponding 𝑅 1.4 values inferred from earlier comprehensive Bayesian analyses of NS observational data, we systematically and rigorously develop the most probable 𝑅 1.4 formulas with varying statistical accuracy and technical complexity. Here, the most important EOS parameters for determining 𝑅 1.4 are found consistently in each of the feature selection processes to be (in order of decreasing importance): curvature 𝐾 sym , slope 𝐿, skewness 𝐽 sym of nuclear symmetry energy, skewness 𝐽 0 , incompressibility 𝐾 0 of symmetric nuclear matter, and the magnitude 𝐸 sym ⁡(𝜌 0 ) of symmetry energy at the saturation density 𝜌 0 of nuclear matter.

Bayesian methods↗

DOE OSTI2021

Attacking the IEC-61131 Logic Engine in Programmable Logic Controllers in Industrial Control Systems

In industrial control systems (ICS), programmable logic controllers (PLCs) directly monitor and control a physical process such as nuclear power plants, gas pipelines, and water treatment. They are equipped with a control logic written in IEC-61131 languages (e.g., ladder logic and structured text) that defines how a PLC should control a physical process. A PLC's control logic is a usual target of a cyberattack to sabotage a physical process. For instance, Stuxnet targets a control logic of a Siemens S7-300 PLC to damage a nuclear facility's centrifuges. The existing attacks in the literature generally focus only on injecting malicious control logic into a PLC. This paper presents a new dimension of control logic attacks that target the control logic engine (responsible for running a control logic) of a PLC. It demonstrates that a cyberattack can disable the control logic engine successfully by exploiting inherent PLC features such as program mode and starting/stopping engine. We develop two novel case studies on control logic engine attacks by employing the MITRE ATT\&CK knowledge base on the real-world PLCs used in industry settings, i.e., 1) Schweitzer Engineering Laboratory (SEL)'s Real-Time Automation Controller (SEL-3505 RTAC) equipped with security features such as encrypted traffic and device-level access control, and 2) traditional PLCs, i.e., Schneider Electric's Modicon M221, Allen-Bradley's MicroLogix 1400 and 1100 that do not have security features. The case studies present the internals of the logic engine attacks and facilitate the ICS research community and industry to understand the attack vectors on the control logic engine. We evaluate the effectiveness of the control engine attacks on a power substation, a 4-floor elevator, and a conveyor belt to demonstrate their real-world impact of halting a physical process.

Ali qasim, Syed↗

DOE OSTI2025

Search for dark matter produced in association with a Higgs boson decaying to a τ lepton pair in proton-proton collisions at $\sqrt{s}=13$ TeV

A search for dark matter particles produced in association with a Higgs boson decaying into a pair of τ leptons is performed using data collected in proton-proton collisions at a center-of-mass energy of 13 TeV with the CMS detector. The analysis is based on a data set corresponding to an integrated luminosity of 101 fb −1 collected in 2017–2018. No significant excess over the expected standard model background is observed. This result is interpreted within the frameworks of the 2HDM+a and baryonic Z′ benchmark simplified models. The 2HDM+a model is a type-II two-Higgs-doublet model featuring a heavy pseudoscalar with an additional light pseudoscalar. Upper limits at 95% confidence level are set on the product of the production cross section and the branching fraction for each of these two simplified models. Heavy pseudoscalar boson masses between 400 and 700 GeV are excluded for a light pseudoscalar mass of 100 GeV. For the baryonic Z′ model, a statistical combination is made with an earlier search based on a data set of 36 fb −1 collected in 2016. In this model, Z′ boson masses up to 1050 GeV are excluded for a dark matter particle mass of 1 GeV.

Dark Matter↗

DOE OSTI2023

Dimensionally reduced machine learning model for predicting single component octanol–water partition coefficients

Abstract MF-LOGP, a new method for determining a single component octanol–water partition coefficients ( $$LogP$$ LogP ) is presented which uses molecular formula as the only input. Octanol–water partition coefficients are useful in many applications, ranging from environmental fate and drug delivery. Currently, partition coefficients are either experimentally measured or predicted as a function of structural fragments, topological descriptors, or thermodynamic properties known or calculated from precise molecular structures. The MF-LOGP method presented here differs from classical methods as it does not require any structural information and uses molecular formula as the sole model input. MF-LOGP is therefore useful for situations in which the structure is unknown or where the use of a low dimensional, easily automatable, and computationally inexpensive calculations is required. MF-LOGP is a random forest algorithm that is trained and tested on 15,377 data points, using 10 features derived from the molecular formula to make $$LogP$$ LogP predictions. Using an independent validation set of 2713 data points, MF-LOGP was found to have an average $$RMSE$$ RMSE = 0.77 ± 0.007, $$MAE$$ MAE = 0.52 ± 0.003, and $${R}^{2}$$ R 2 = 0.83 ± 0.003. This performance fell within the spectrum of performances reported in the published literature for conventional higher dimensional models ( $$RMSE$$ RMSE = 0.42–1.54, $$MAE$$ MAE = 0.09–1.07, and $${R}^{2}$$ R 2 = 0.32–0.95). Compared with existing models, MF-LOGP requires a maximum of ten features and no structural information, thereby providing a practical and yet predictive tool. The development of MF-LOGP provides the groundwork for development of more physical prediction models leveraging big data analytical methods or complex multicomponent mixtures. Graphical Abstract

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

DOE OSTI2026

MATEY: multiscale adaptive transformer models for spatiotemporal physical systems

Accurate representation of the multiscale features in spatiotemporal physical systems using vision transformer architectures requires extremely long, computationally prohibitive token sequences. To address this issue, we propose two novel adaptive tokenization schemes that dynamically adjust patch sizes based on local features: one ensures convergent behavior to uniform patch refinement, while the other offers better computational efficiency. Moreover, we present a set of spatiotemporal attention schemes, where the temporal or axial spatial dimensions are decoupled, to evaluate their baseline computational and data efficiencies and to determine whether adaptive tokenization can improve this performance. We assess the performance of the proposed multiscale adaptive model, MATEY, in a sequence of experiments. Compared to a full spatiotemporal attention scheme or a scheme that decouples only the temporal dimension, we find that fully decoupled axial attention is less efficient and expressive, requiring more training time and model parameters to achieve the same accuracy. The experiments on the adaptive tokenization schemes show that, compared to a uniformly refined model, the proposed schemes achieve comparable or improved accuracy at a much lower cost in the tested two-dimensional settings. While the asymptotic analysis suggests the potential for favorable scaling, empirical validation at substantially longer sequence lengths remains to be performed in future work. Finally, we demonstrate in two fine-tuning tasks featuring different physics that models pretrained on PDEBench data outperform the ones trained from scratch, especially in the low data regime with frozen attention.

adaptive tokenization↗