Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “temporal computing”

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 109 records · Page 6
All sourcesNASA NTRSDOE OSTI
DOE OSTI2018

Coarse-Grained Double-Stranded RNA Model from Quantum-Mechanical Calculations

A coarse-grained model for simulating structural properties of double-stranded RNA is developed with parameters obtained from quantum-mechanical calculations. This model follows previous parametrization for double stranded DNA, which is based on mapping the all-atom picture to a coarse-grained four-bead scheme. Chemical and structural differences between RNA and DNA have been taken into account for the model development. The parametrization is based on simulations using density functional theory (DFT) on separate units of the RNA molecule without implementing experimental data. The total energy is decomposed into four terms of physical significance: hydrogen bonding interaction, stacking interactions, backbone interactions, and electrostatic interactions. The first three interactions are treated within DFT, whereas the last one is included within a mean field approximation. Our double-stranded RNA coarse-grained model predicts stable helical structures for RNA. Other characteristics, such as structural or mechanical properties are reproduced with a very good accuracy. The development of the coarse-grained model for RNA allows extending the spatial and temporal length scales accessed by computer simulations and being able to model RNA-related biophysical processes, as well as novel RNA nanostructures.

Cruz-León, Sergio↗
DOE OSTI2025

A Comparative Study of Physics‐Informed and Data‐Driven Neural Networks for Compound Flood Simulation at River‐Ocean Interfaces: A Case Study of Hurricane Irene

Simulating compound flooding (CF) at the river-ocean interface within large-scale Earth System Models (ESMs) presents significant challenges due to complex interactions between river discharge, storm surge, and tides. This study assesses the comparative advantages of physics-informed and data-driven machine learning (ML) approaches for enhancing local ESM performance. We systematically compare data-driven neural network models (i.e., CNNs, U-Net, Long Short-Term Memory (LSTM), Gated Recurrent Unit), and physics-informed neural network (PINN) models, including vanilla PINN and a finite-difference-based PINN (FD-PINN). Specifically, FD-PINN is introduced to enhance computational efficiency, accelerating vanilla PINNs by ∼6.5 times while improving accuracy. To enhance data-driven model training, a new data-generation approach is developed to sample historical fluvial and coastal flood events, which ensures a robust data set for extreme event prediction. The models are evaluated using a realistic one-dimensional river domain extracted from an ESM's river mesh and the Hurricane Irene event as an independent test case. Results show that FD-PINN achieves accurate predictions with significantly reduced computational costs relative to vanilla PINNs. Among data-driven models, the best overall performance is achieved by a CNN-LSTM hybrid, which balances accuracy and efficiency. While a fully connected CNN (CNN-FC) provides the best accuracy, it incurs high computational cost. Architectures lacking strong temporal modeling tend to underperform on unseen events. These findings highlight the importance of sequence-aware designs for robust generalization. This study reveals the trade-offs between physics-informed and data-driven models and proposes an adaptive hybrid framework for integrating ML into ESMs to enhance local flood simulations.

Earth Systems Modeling↗
DOE OSTI2018

Polarimetric Interferometric SAR Change Detection Discrimination

A coherent change detection (CCD) image, computed from a geometrically matched, temporally separated pair of complex-valued synthetic aperture radar (SAR) image sets, conveys the pixel-level equivalence between the two observations. Low-coherence values in a CCD image are typically due to either some physical change in the corresponding pixels or a low signal-to-noise observation. A CCD image does not directly convey the nature of the change that occurred to cause low coherence. In this paper, we introduce a mathematical framework for discriminating between different types of change within a CCD image. We utilize the extra degrees of freedom and information from polarimetric interferometric SAR (PolInSAR) data and PolInSAR processing techniques to define a 29-dimensional feature vector that contains information capable of discriminating between different types of change in a scene. We also propose two change-type discrimination functions that can be trained with feature vector training data and demonstrate change-type discrimination on an example image set for three different types of change. In conclusion, we also describe and characterize the performance of the two proposed change-type discrimination functions by way of receiver operating characteristic curves, confusion matrices, and pass matrices.

58 GEOSCIENCES↗
DOE OSTI2025

Hydrological connectivity: a review and emerging strategies for integrating measurement, modeling, and management

This review synthesizes methods for measuring, modeling, and managing hydrologic connectivity, offering pathways to improve practices and address environmental challenges (e.g., climate change) and sustainability. As a key driver of water movement and nutrient cycling, hydrologic connectivity influences flood mitigation, water quality regulation, and biodiversity conservation. However, traditional field-based methods (e.g., dye tracing), indirect measurements (e.g., runoff analysis), and remote sensing techniques (e.g., InSAR) often struggle to capture the complexity of catchment-scale interactions. Similarly, modeling approaches—including process-based and percolation theory-based models, graph theory, and entropy-based metrics—face limitations in fully representing these interconnected processes. Both modeling and measurement techniques are constrained by inadequate spatial and temporal coverage, high data demands, computational complexity, and difficulties in representing subsurface connectivity. Subsequently, we critique current management practices that prioritize isolated variables (e.g., streamflow, sediment transport) over system-wide strategies and emphasize the need for adaptive, connectivity-based approaches in water resource planning and restoration. Moving forward, we highlight the importance of interdisciplinary collaboration, technological innovations (e.g., AI-driven modeling, real-time monitoring), and integrated frameworks to improve connectivity measurement, modeling, and adaptive management to restore fragmented hydrologic networks. This integrated approach sets the stage for transformative water resource management, fostering proactive policy development and stakeholder engagement.

Dwivedi, Dipankar↗
DOE OSTI2018

Relativistic Hydrodynamics with Wavelets

Methods to solve the relativistic hydrodynamic equations are important in a large number of astrophysical simulations, which may be very dynamic and involve multiscale features. This requires computational methods that are highly adaptive and capable of automatically resolving numerous localized features and instabilities that emerge across the computational domain and over many temporal scales. While this has been historically accomplished with adaptive-mesh-refinement-based methods, alternatives using wavelet bases and the wavelet transformation have recently achieved significant success in adaptive representation for advanced engineering applications. The current work presents a new method, extending the wavelet adaptive multiresolution representation method, for the integration of the relativistic hydrodynamic equations using iterated interpolating wavelets and introduces a highly adaptive implementation for multidimensional simulation. We note the wavelet coefficients provide a direct measure of the local approximation error for the solution and place collocation points that naturally adapt to the fluid flow while providing good conservation of fluid quantities. The resulting implementation, OAHU , is applied to a series of demanding 1D and 2D problems that explore high Lorentz factor outflows and the formation of several instabilities, including the Kelvin–Helmholtz instability and the Rayleigh–Taylor instability.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2026

Integrating AI Data Centers with the Power Grid

The rapid expansion of artificial intelligence (AI) has triggered an unprecedented surge in electricity demand, with US data center energy use projected to double or triple 2023 levels by 2028. This exponential growth places strain on grid infrastructure, which can hinder timely construction of desired computing capacity. To bridge this supply-demand gap, utilities and AI developers are increasingly turning to demand flexibility, a strategy that incentivizes shifting or reducing power use during peak periods of grid stress. Data centers are uniquely equipped for flexible operations due to their digital workloads, built-in redundancy, and onsite energy assets. This article outlines four primary mechanisms to enable data center flexibility: computational load flexibility (shifting tasks temporally or geographically), flexible use of core facility infrastructure adjustments, energy storage utilization, and onsite electricity generation. To encourage adoption, utilities are deploying new tariff designs, including voluntary interruptible service riders, mandated flexibility requirements, and streamlined interconnection processes for flexible loads. For the highly capitalized and rapidly growing AI industry, the primary motivators for embracing these strategies are expediting facility interconnection, satisfying emerging regulatory mandates, and mitigating community resistance. While demand flexibility cannot substitute the long-term need for new bulk power generation, it serves as an essential, immediate solution for enabling near-term deployment. By transforming data centers from grid stressors into stabilizing assets, flexible operations can ensure reliable grid integration, ease market pressures, and support a resilient power system.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2019

Nonnegative tensor decomposition with custom clustering for microphase separation of block copolymers

High–dimensional datasets are becoming ubiquitous in many applications and therefore unsupervised tensor methods to interrogate them are needed. Here, we report a new unsupervised machine learning (ML) approach (NTFk) based on nonnegative tensor factorization integrated with a custom k–means clustering. We demonstrate the ability of NTFk to extracting temporal and spatial features of phase separation of copolymers as they are modeled by self–consistent field theory. Microphase separation of block copolymers has been extensively studied both experimentally and theoretically. However, the interpretation of computer simulations and/or experimental data, representing temporal and spatial changes of molecular species concentration is still a challenging task. Thus, extracting the phase diagram from simulations or experimental data as well as the interpretation of data requires discernment of the model/experimental parameters (such as, temperature, concentrations, the number of molecular species and the interaction between species) impact on the microphase separation process. An attractive and unique aspect of the introduced ML method is that it ensures the nonnegativity of the extracted latent features. Nonnegativity is an essential constraint needed to obtain interpretable and sparse latent features that are parts–based representation of the data. The custom clustering in NTFk serves to estimate the number of latent features in the data.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Application of a temporal multiscale method for efficient simulation of degradation in PEM Water Electrolysis under dynamic operating conditions

Hydrogen is emerging as a vital energy carrier, driven by the need to reduce carbon emissions. Proton Electrolyte Membrane Water Electrolysis (PEMWE) enables hydrogen production under fluctuating renewable power conditions but requires improved understanding and stability of the anode catalyst layer under dynamic operating conditions, especially with low noble metal loadings. Long-term degradation experiments are both time-consuming and costly; therefore, a systematic, model-aided approach is essential. In the present work, a temporal multiscale method is applied to reduce the computational effort of simulating long-term degradation processes in PEMWE, with an exemplary focus on catalyst dissolution. A mechanistic model incorporating the oxygen evolution reaction, catalyst dissolution, and hydrogen permeation from the cathode to the anode was hypothesized and implemented. In this way, the local periodicity of transport and reaction processes in dynamic PEMWE operation, which influence the gradual degradation of the catalyst layer, is captured. The temporal multiscale method significantly reduces the computational effort of simulation, decreasing processing time from hours to mere minutes. This efficiency gain is attributed to the limited evolution of Slow-Scale variables during each period of time P of the Fast-Scale variables. Consequently, simulation is required only until local periodicity is achieved within each Slow-Scale time step. Hence, the fully resolved dynamic problem is decoupled into these two scales, employing a heterogeneous multiscale technique. The developed approach effectively accelerates parameter estimation and predictive simulations, supporting systematic modeling of PEMWE degradation under dynamic conditions.

08 HYDROGEN↗
DOE OSTI2021

A multiresolution adaptive wavelet method for nonlinear partial differential equations

We report the multiscale complexity of modern problems in computational science and engineering can prohibit the use of traditional numerical methods in multi-dimensional simulations. Therefore, novel algorithms are required in these situations to solve partial differential equations (PDEs) with features evolving on a wide range of spatial and temporal scales. To meet these challenges, we present a multiresolution wavelet algorithm to solve PDEs with significant data compression and explicit error control. We discretize in space by projecting fields and spatial derivative operators onto wavelet basis functions. We provide error estimates for the wavelet representation of fields and their derivatives. Then, our estimates are used to construct a sparse multiresolution discretization which guarantees the prescribed accuracy. Additionally, we embed a predictor-corrector procedure within the temporal integration to dynamically adapt the computational grid and maintain the accuracy of the solution of the PDE as it evolves. We present examples to highlight the accuracy and adaptivity of our approach.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

SENTRA: A Modular Computational Graph Framework for Critical Mineral and Materials Supply Chains: Part I: Network Construction Latent-Quantity Estimation, and Temporal Graph Forecasting

Global supply chains for critical minerals and materials are complex, evolving networks of countries, products, production stages, and trade relationships. Existing analytical approaches are limited by fragmented data and static network representations that do not capture the dynamic production dependencies linking raw materials, intermediate products, and final goods across multiple countries. Trade and production statistics provide only a partial view of domestic production, inventories, and material flows, making it difficult to identify indirect sourcing pathways, hidden dependencies, and embedded foreign exposures. This paper introduces the Supply Chain Exposure Network Tracking and Risk Assessment (SENTRA) framework, a modular graph-based computational framework for constructing, analyzing, and forecasting dynamic supply chain networks. As the first paper in a three-part methodological series, it establishes the computational foundation of SENTRA by constructing a temporal attributed multi-relational graph whose nodes represent product–country pairs and whose edges encode observed trade and within-country value-chain relationships. Statistical estimation and constrained optimization recover latent production, final demand, and product input dependency coefficients while enforcing economic accounting constraints. Graph-derived exposure measures quantify direct, transshipment, value-chain, and multi-hop supply chain dependencies independently of the forecasting model. A temporal graph forecasting architecture based on a relational graph neural network then forecasts the evolution of the graph under mass-balance constraints with distribution-free conformal uncertainty quantification. Validation on the global aluminum supply chain shows that the learned graph representations recover economically meaningful supply chain structure, accurately forecast out-of-sample trade relationships, and produce well-calibrated prediction intervals. Subsequent papers apply this computational foundation to exposure assessment, disruption analysis, and scenario-based policy analysis, and extend the framework to multimaterial supply chain modeling and decision support.

36 MATERIALS SCIENCE↗
DOE OSTI2019

Phonetic acquisition in cortical dynamics, a computational approach

Many computational theories have been developed to improve artificial phonetic classification performance from linguistic auditory streams. However, less attention has been given to psycholinguistic data and neurophysiological features recently found in cortical tissue. We focus on a context in which basic linguistic units-such as phonemes-are extracted and robustly classified by humans and other animals from complex acoustic streams in speech data. We are especially motivated by the fact that 8-month-old human infants can accomplish segmentation of words from fluent audio streams based exclusively on the statistical relationships between neighboring speech sounds without any kind of supervision. In this paper, we introduce a biologically inspired and fully unsupervised neurocomputational approach that incorporates key neurophysiological and anatomical cortical properties, including columnar organization, spontaneous micro-columnar formation, adaptation to contextual activations and Sparse Distributed Representations (SDRs) produced by means of partial N-Methyl-D-aspartic acid (NMDA) depolarization. Its feature abstraction capabilities show promising phonetic invariance and generalization attributes. Our model improves the performance of a Support Vector Machine (SVM) classifier for monosyllabic, disyllabic and trisyllabic word classification tasks in the presence of environmental disturbances such as white noise, reverberation, and pitch and voice variations. Furthermore, our approach emphasizes potential self-organizing cortical principles achieving improvement without any kind of optimization guidance which could minimize hypothetical loss functions by means of-for example-backpropagation. Thus, our computational model outperforms multiresolution spectro-temporal auditory feature representations using only the statistical sequential structure immerse in the phonotactic rules of the input stream.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2017

PlantCV v2: Image analysis software for high-throughput plant phenotyping

Systems for collecting image data in conjunction with computer vision techniques are a powerful tool for increasing the temporal resolution at which plant phenotypes can be measured non-destructively. Computational tools that are flexible and extendable are needed to address the diversity of plant phenotyping problems. We previously described the Plant Computer Vision (PlantCV) software package, which is an image processing toolkit for plant phenotyping analysis. The goal of the PlantCV project is to develop a set of modular, reusable, and repurposable tools for plant image analysis that are open-source and community-developed. Here in this paper we present the details and rationale for major developments in the second major release of PlantCV. In addition to overall improvements in the organization of the PlantCV project, new functionality includes a set of new image processing and normalization tools, support for analyzing images that include multiple plants, leaf segmentation, landmark identification tools for morphometrics, and modules for machine learning.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2018

Unraveling the surface chemistry processes in lithiated and boronized plasma material interfaces under extreme conditions

The review of recent theoretical and experimental research on the complex surface chemistry processes that evolve from low-Z material conditioning on plasma-facing materials under extreme fusion plasma conditions is presented. A combination of multi-scale computational physics and chemistry modeling with real-time diagnosis of the plasma-material interface in tokamak fusion plasma edge is complemented by ex-vessel in-situ single-effect experimental facilities to unravel the evolving characteristics of low-Z components under irradiation. Effects of the lithium and boron coatings at carbon surfaces to the retention of deuterium and chemical sputtering of the plasma-facing surfaces are discussed in detail. The critical role of oxygen in the surface chemistry during hydrogen-fuel irradiation is found to drive the kinetics and dynamics of these surfaces as they interact with fusion edge plasma that ultimately could have profound effects on fusion plasma confinement behavior. Computational studies also extend in spatio-temporal scales not accessible by empirical means and therefore open the opportunity for a strategic approach at irradiation surface science studies that combined these powerful computational tools with in-vessel and ex-vessel in-situ diagnostics.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2023

A Novel Spatial-Temporal Variational Quantum Circuit to Enable Deep Learning on NISQ Devices

Quantum computing presents a promising approach for machine learning with its capability for extremely parallel computation in high-dimension through superposition and entanglement. Despite its potential, existing quantum learning algorithms, such as Variational Quantum Circuits (VQCs), face challenges in handling more complex datasets, particularly those that are not linearly separable. What’s more, it encounters the deployability issue, making the learning models suffer a drastic accuracy drop after deploying them to the actual quantum devices. To overcome these limitations, this paper proposes a novel spatial-temporal design, namely “ST-VQC”, to integrate nonlinearity in quantum learning and improve the robustness of the learning model to noise. Specifically, ST-VQC can extract spatial features via a novel block-based encoding quantum sub-circuit coupled with a layer-wise computation quantum sub-circuit to enable temporal-wise deep learning. Additionally, a SWAP-Free physical circuit design is devised to improve robustness. These designs bring a number of hyperparameters. After a systematic analysis of the design space for each design component, an automated optimization framework is proposed to generate the ST-VQC quantum circuit. The proposed ST-VQC has been evaluated on two IBM quantum processors, ibm-cairo with 27 qubits and ibmq-lima with 7 qubits to assess its effectiveness. The results of the evaluation on the standard dataset for binary classification show that ST-VQC can achieve over 30% accuracy improvement compared with existing VQCs on actual quantum computers. Moreover, on a non-linear synthetic dataset, the STVQC outperforms a linear classifier by 27.9%, while the linear classifier using classical computing outperforms the existing VQC by 15.58%.

Li, Jinyang↗
DOE OSTI2022

Neural Network for Principle of Least Action

The principle of least action is the cornerstone of classical mechanics, theory of relativity, quantum mechanics, and thermodynamics. Here, we describe how a neural network (NN) learns to find the trajectory for a Lennard-Jones (LJ) system that maintains balance in minimizing the Onsager–Machlup (OM) action and maintaining the energy conservation. The phase-space trajectory thus calculated is in excellent agreement with the corresponding results from the “ground-truth” molecular dynamics (MD) simulation. Furthermore, we show that the NN can easily find structural transformation pathways for LJ clusters, for example, the basin-hopping transformation of an LJ 38 from an incomplete Mackay icosahedron to a truncated face-centered cubic octahedron. Unlike MD, the NN computes atomic trajectories over the entire temporal domain in one fell swoop, and the NN time step is a factor of 20 larger than the MD time step. The NN approach to OM action is quite general and can be adapted to model morphometrics in a variety of applications.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2018

Mode penetration induced by transient magnetic perturbations

Three dimensional magnetic fields in tokamaks can induce forced magnetic reconnection (FMR) and produce magnetic islands on resonant surfaces. Conventional analytic solutions to FMR focus on describing the time asymptotic state given a steady-state field error. The purpose of this report is to understand the nonlinear dynamics of mode penetration, an evolution from a high-slip, flow-screened metastable equilibrium into a low-slip, field-penetrated metastable equilibrium. Here, we extend previous work by incorporating a temporally varying external magnetic field as a simple model for a magnetohydrodynamic (MHD) event that produces resonant magnetic perturbations. Proof-of-principle, extended-MHD, NIMROD computations vary parameterizations of the transient external perturbation to probe the threshold for mode penetration. We test these computational results against analytical theory that captures the temporal evolution properties of the electromagnetic and viscous forces during and after a transient. We find qualitative agreement between computational and analytical results. However, computational tools are necessary to accurately capture the threshold conditions for mode penetration induced by an MHD transient.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2018

Experimentally Validated Numerical Models of Non-Isothermal Turbulent Mixing in High Temperature Reactors

In the Very High Temperature Reactor (VHTR), the non-uniform heat generation and flow maldistribution of coolant in the core generates conditions leading into the lower plenum with many jets at different temperatures and velocities. The hope is that the combined flow becomes fully mixed in a short distance so that the fluctuating temperatures in the mixing region do not cause unwanted thermal stresses in the supporting structures. These cyclic thermal stresses can cause degradation of the wall microstructure including crack initiation and growth, ultimately leading to the fatigue failure in the lower plenum, and can be accelerated by the high temperature environment in the VHTR. The focus of this project was to provide experimental data for future validation of computational efforts aimed at capturing the turbulent mixing in the lower plenum. Computational fluid dynamics (CFD) is also used as part of this effort to assess the requirements to model fundamental flows with application to the lower plenum as well at the applied flow conditions that mimic the experimental facilities. These efforts will lay the groundwork for future incorporation of stress analysis as well as fatigue and failure studies for multiple VHTR components. An additional emphasis is placed on reduced order models (ROM) using canonical flows, which will help to balance computational expense and accuracy and enable future work in uncertainty quantification. The flow in the lower plenum is extremely complex with hundreds of cooling jets entering and then being forced to change directions by 90° while navigating through an array of cylindrical support posts. Generating experimental data for such a complicated flow is difficult, not to mention that the computational requirements to generate an adequate temporal and spatial sampling is extremely expensive and time consuming. Therefore, the ultimate focus of this research is a “unit cell” which consists of six jets in the presence of a crossflow, and seven cylindrical support posts. It is scaled to achieve similar ranges of key dimensionless numbers. This serves to mimic flow conditions in different regions of the lower plenum by adjusting velocities, temperatures, and orientations, and represents a compromise between the complexity inside the full-scale application and the ideal conditions of a more fundamental study. In addition to the unit cell experiments and simulations, the complex flows in the VHTR lower plenum are broken down into a number of simplified (in some cases even fundamental) flows, two of which are turbulent jets and flow past cylinders. Each of these are studied, computationally and experimentally for isothermal and non-isothermal conditions. Flows where trusted experimental data is available in the literature (e.g., turbulent free jet), additional experiments are not needed. The result of these efforts is a database of experimental data and multiple computational studies, which provide an outlook to the requirements, needed to capture turbulent mixing flow physics The ROM results provide assessments for appropriate resource-light models that balance accuracy with reduced computational burdens.

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗
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↗