Search

We introduce a novel method to enhance cross-language code translation from Fortran to C++ by integrating task-specific embedding alignment into a Retrieval-Augmented Generation (RAG) framework. Unlike conventional retrieval approaches that utilize generic embeddings agnostic to the downstream task, our strategy aligns the retrieval model directly with the objective of maximizing translation quality, as quantified by the CodeBLEU metric. This alignment ensures that the embeddings are semantically and syntactically meaningful for the specific code translation task. Our methodology involves constructing a dataset of 25,000 Fortran code snippets sourced from Stack-V2 dataset and generating their corresponding C++ translations using the LLaMA 3.1-8B language model. We compute pairwise CodeBLEU scores between the generated translations and ground truth examples to capture fine-grained similarities. These scores serve as supervision signals in a contrastive learning framework, where we optimize the embedding model to retrieve Fortran-C++ pairs that are most beneficial for improving the language model's translation performance. By integrating these CodeBLEU-optimized embeddings into the RAG framework, our approach significantly enhances both retrieval accuracy and code generation quality over methods employing generic embeddings. On the HPC Fortran2C++ dataset, our method elevates the average CodeBLEU score from 0.64 to 0.73, achieving a 14% relative improvement. On the Numerical Recipes dataset, we observe an increase from 0.52 to 0.60, marking a 15% relative improvement. Importantly, these gains are realized without any fine-tuning of the language model, underscoring the efficiency and practicality of our approach.

For decades, corporations and governments have relied on scanned documents to record vast amounts of information. However, extracting this information is a slow and tedious process due to the overwhelming amount of documents. The rise of vision language models presents a way to efficiently and accurately extract the information out of these documents. The current automated workflow often requires a two-step approach involving the extraction of information using optical character recognition software, and subsequent usage of large language models for processing this information. Unfortunately, these methods encounter significant challenges when dealing with noisy scanned documents. The high information density of such documents often necessitates using computationally expensive language models to effectively reduce noise. In this study, we propose PatchFinder, an algorithm that builds upon Vision Language Models (VLMs) to address the information extraction task. First, we devise a confidence-based score, called Patch Confidence, based on the Maximum Softmax Probability of the VLMs' output to measure the model's confidence in its predictions. Then, PatchFinder utilizes that score to determine a suitable patch size, partition the input document into overlapping patches of that size, and generate confidence-based predictions for the target information. Our experimental results show that PatchFinder can leverage Phi-3v, a 4.2 billion parameter vision language model, to achieve an accuracy of 94% on our dataset of 190 noisy scanned documents, surpassing the performance of ChatGPT-4o by 18.5 percentage points., Comment: This paper has been accepted to IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2025

We investigate the use of the Senseiver, a transformer neural network designed for sparse sensing applications, to estimate full-field surface height measurements of tsunami waves from sparse observations. The model is trained on a large ensemble of simulated data generated via a shallow water equations solver, which we show to be a faithful reproduction for the underlying dynamics by comparison to historical events. We train the model on a dataset consisting of 8 tsunami simulations whose epicenters correspond to historical USGS earthquake records, and where the model inputs are restricted to measurements obtained at actively deployed buoy locations. We test the Senseiver on a dataset consisting of 8 simulations not included in training, demonstrating its capability for extrapolation. The results show remarkable resolution of fine scale phase and amplitude features from the true field, provided that at least a few of the sensors have obtained a non-zero signal. Throughout, we discuss which forecasting techniques can be improved by this method, and suggest ways in which the flexibility of the architecture can be leveraged to incorporate arbitrary remote sensing data (eg. HF Radar and satellite measurements) as well as investigate optimal sensor placements.

Understanding material failure is critical for designing stronger and lighter structures by identifying weaknesses that could be mitigated. Existing full-physics numerical simulation techniques involve trade-offs between speed, accuracy, and the ability to handle complex features like varying boundary conditions, grid types, resolution, and physical models. We present the first foundation model specifically designed for predicting material failure, leveraging large-scale datasets and a high parameter count (up to 3B) to significantly improve the accuracy of failure predictions. In addition, a large language model provides rich context embeddings, enabling our model to make predictions across a diverse range of conditions. Unlike traditional machine learning models, which are often tailored to specific systems or limited to narrow simulation conditions, our foundation model is designed to generalize across different materials and simulators. This flexibility enables the model to handle a range of material properties and conditions, providing accurate predictions without the need for retraining or adjustments for each specific case. Our model is capable of accommodating diverse input formats, such as images and varying simulation conditions, and producing a range of outputs, from simulation results to effective properties. It supports both Cartesian and unstructured grids, with design choices that allow for seamless updates and extensions as new data and requirements emerge. Our results show that increasing the scale of the model leads to significant performance gains (loss scales as $N^{-1.6}$, compared to language models which often scale as $N^{-0.5}$)., Comment: Accepted at NeurIPS 2024 "Foundation Models for Science: Progress, Opportunities, and Challenges" Workshop

This paper presents a novel benchmark where the large language model (LLM) must write code that computes integer sequences from the Online Encyclopedia of Integer Sequences (OEIS), a widely-used resource for mathematical sequences. The benchmark is designed to evaluate both the correctness of the generated code and its computational efficiency. Our benchmark reveals that the o1 series of models outperform other frontier models from OpenAI, Anthropic, Meta, and Google in accuracy and cheating rates across both easy and hard integer sequences. In order to ensure models do not exploit memorized sequence values, we introduce an automated cheating detection mechanism that flags the use of lookup tables and validated this automation against human cheating evaluations. This benchmark provides a meaningful challenge for current LLMs, offering insights into their mathematical reasoning and code writing capabilities, which can guide future research directions and model development in mathematical reasoning and code synthesis.

The advent of large language models (LLMs) has significantly advanced the field of code translation, enabling automated translation between programming languages. However, these models often struggle with complex translation tasks due to inadequate contextual understanding. This paper introduces a novel approach that enhances code translation through Few-Shot Learning, augmented with retrieval-based techniques. By leveraging a repository of existing code translations, we dynamically retrieve the most relevant examples to guide the model in translating new code segments. Our method, based on Retrieval-Augmented Generation (RAG), substantially improves translation quality by providing contextual examples from which the model can learn in real-time. We selected RAG over traditional fine-tuning methods due to its ability to utilize existing codebases or a locally stored corpus of code, which allows for dynamic adaptation to diverse translation tasks without extensive retraining. Extensive experiments on diverse datasets with open LLM models such as Starcoder, Llama3-70B Instruct, CodeLlama-34B Instruct, Granite-34B Code Instruct, and Mixtral-8x22B, as well as commercial LLM models like GPT-3.5 Turbo and GPT-4o, demonstrate our approach's superiority over traditional zero-shot methods, especially in translating between Fortran and CPP. We also explored varying numbers of shots i.e. examples provided during inference, specifically 1, 2, and 3 shots and different embedding models for RAG, including Nomic-Embed, Starencoder, and CodeBERT, to assess the robustness and effectiveness of our approach., Comment: LLM for code translation

This study introduces a hybrid fluid simulation approach that integrates generative diffusion models with physics-based simulations, aiming at reducing the computational costs of flow simulations while still honoring all the physical properties of interest. These simulations enhance our understanding of applications such as assessing hydrogen and CO$_2$ storage efficiency in underground reservoirs. Nevertheless, they are computationally expensive and the presence of nonunique solutions can require multiple simulations within a single geometry. To overcome the computational cost hurdle, we propose a hybrid method that couples generative diffusion models and physics-based modeling. We introduce a system to condition the diffusion model with a geometry of interest, allowing to produce variable fluid saturations in the same geometry. While training the model, we simultaneously generate initial conditions and perform physics-based simulations using these conditions. This integrated approach enables us to receive real-time feedback on a single compute node equipped with both CPUs and GPUs. By efficiently managing these processes within one compute node, we can continuously evaluate performance and stop training when the desired criteria are met. To test our model, we generate realizations in a real Berea sandstone fracture which shows that our technique is up to 4.4 times faster than commonly used flow simulation initializations.

Recreating complex, high-dimensional global fields from limited data points is a grand challenge across various scientific and industrial domains. Given the prohibitive costs of specialized sensors and the frequent inaccessibility of certain regions of the domain, achieving full field coverage is typically not feasible. Therefore, the development of algorithms that intelligently improve sensor placement is of significant value. In this study, we introduce a general approach that employs differentiable programming to exploit sensor placement within the training of a neural network model in order to improve field reconstruction. We evaluated our method using two distinct datasets; the results show that our approach improved test scores. Ultimately, our method of differentiable placement strategies has the potential to significantly increase data collection efficiency, enable more thorough area coverage, and reduce redundancy in sensor deployment.

Pore-scale simulations accurately describe transport properties of fluids in the subsurface. These simulations enhance our understanding of applications such as assessing hydrogen storage efficiency and forecasting CO$_2$ sequestration processes in underground reservoirs. Nevertheless, they are computationally expensive due to their mesoscopic nature. In addition, their stationary solutions are not guaranteed to be unique, so multiple runs with different initial conditions must be performed to ensure sufficient sample coverage. These factors complicate the task of obtaining representative and reliable forecasts. To overcome the high computational cost hurdle, we propose a hybrid method that couples generative diffusion models and physics-based modeling. Upon training a generative model, we synthesize samples that serve as the initial conditions for physics-based simulations. We measure the relaxation time (to stationary solutions) of the simulations, which serves as a validation metric and early-stopping criterion. Our numerical experiments revealed that the hybrid method exhibits a speed-up of up to 8.2 times compared to commonly used initialization methods. This finding offers compelling initial support that the proposed diffusion model-based hybrid scheme has potentials to significantly decrease the time required for convergence of numerical simulations without compromising the physical robustness.

Modeling effective transport properties of 3D porous media, such as permeability, at multiple scales is challenging as a result of the combined complexity of the pore structures and fluid physics - in particular, confinement effects which vary across the nanoscale to the microscale. While numerical simulation is possible, the computational cost is prohibitive for realistic domains, which are large and complex. Although machine learning models have been proposed to circumvent simulation, none so far has simultaneously accounted for heterogeneous 3D structures, fluid confinement effects, and multiple simulation resolutions. By utilizing numerous computer science techniques to improve the scalability of training, we have for the first time developed a general flow model that accounts for the pore-structure and corresponding physical phenomena at scales from Angstrom to the micrometer. Using synthetic computational domains for training, our machine learning model exhibits strong performance (R$^2$=0.9) when tested on extremely diverse real domains at multiple scales.

In this paper, we introduce Pysimfrac, a open-source python library for generating 3-D synthetic fracture realizations, integrating with fluid simulators, and performing analysis. Pysimfrac allows the user to specify one of three fracture generation techniques (Box, Gaussian, or Spectral) and perform statistical analysis including the autocorrelation, moments, and probability density functions of the fracture surfaces and aperture. This analysis and accessibility of a python library allows the user to create realistic fracture realizations and vary properties of interest. In addition, Pysimfrac includes integration examples to two different pore-scale simulators and the discrete fracture network simulator, dfnWorks. The capabilities developed in this work provides opportunity for quick and smooth adoption and implementation by the wider scientific community for accurate characterization of fluid transport in geologic media. We present Pysimfrac along with integration examples and discuss the ability to extend Pysimfrac from a single complex fracture to complex fracture networks.

We model flow and transport in three-dimensional fracture networks with varying degrees of fracture-to-fracture aperture/permeability heterogeneity and network density to show how changes in these properties can cause the emergence of anomalous flow and transport behavior. If fracture-to-fracture aperture heterogeneity is increased in sparse networks, velocity fluctuations can inhibit high flow rates and solute transport can be delayed, even in cases where hydraulic aperture is monotonically increased. As the density of the networks is increased, more connected pathways allow for particles to bypass these effects. We discover transition behavior where with relatively few connected pathways in a network from inflow to outflow boundaries, the first arrival times of particles are not heavily affected by fracture-to-fracture aperture heterogeneity, but the scaling behavior of the tails is strongly influenced due to the particles being forced to sample some of the heterogeneity in the velocity field caused by aperture differences. These results reinforce the importance of considering multi-scale effects in fractured systems and can inform flow and transport processes in both natural and engineered fracture systems, especially the latter where high aperture fractures are often stimulated and connect to existing fracture networks with smaller apertures.

Typical generative diffusion models rely on a Gaussian diffusion process for training the backward transformations, which can then be used to generate samples from Gaussian noise. However, real world data often takes place in discrete-state spaces, including many scientific applications. Here, we develop a theoretical formulation for arbitrary discrete-state Markov processes in the forward diffusion process using exact (as opposed to variational) analysis. We relate the theory to the existing continuous-state Gaussian diffusion as well as other approaches to discrete diffusion, and identify the corresponding reverse-time stochastic process and score function in the continuous-time setting, and the reverse-time mapping in the discrete-time setting. As an example of this framework, we introduce ``Blackout Diffusion'', which learns to produce samples from an empty image instead of from noise. Numerical experiments on the CIFAR-10, Binarized MNIST, and CelebA datasets confirm the feasibility of our approach. Generalizing from specific (Gaussian) forward processes to discrete-state processes without a variational approximation sheds light on how to interpret diffusion models, which we discuss., Comment: 29 pages, 13 figures, 2 tables. Accepted by the 40th International Conference on Machine Learning, Honolulu, Hawaii, USA

Spatial nonstationarity, the location variance of features' statistical distributions, is ubiquitous in many natural settings. For example, in geological reservoirs rock matrix porosity varies vertically due to geomechanical compaction trends, in mineral deposits grades vary due to sedimentation and concentration processes, in hydrology rainfall varies due to the atmosphere and topography interactions, and in metallurgy crystalline structures vary due to differential cooling. Conventional geostatistical modeling workflows rely on the assumption of stationarity to be able to model spatial features for the geostatistical inference. Nevertheless, this is often not a realistic assumption when dealing with nonstationary spatial data and this has motivated a variety of nonstationary spatial modeling workflows such as trend and residual decomposition, cosimulation with secondary features, and spatial segmentation and independent modeling over stationary subdomains. The advent of deep learning technologies has enabled new workflows for modeling spatial relationships. However, there is a paucity of demonstrated best practice and general guidance on mitigation of spatial nonstationarity with deep learning in the geospatial context. We demonstrate the impact of two common types of geostatistical spatial nonstationarity on deep learning model prediction performance and propose the mitigation of such impacts using self-attention (vision transformer) models. We demonstrate the utility of vision transformers for the mitigation of nonstationarity with relative errors as low as 10%, exceeding the performance of alternative deep learning methods such as convolutional neural networks. We establish best practice by demonstrating the ability of self-attention networks for modeling large-scale spatial relationships in the presence of commonly observed geospatial nonstationarity.

Throughout computational science, there is a growing need to utilize the continual improvements in raw computational horsepower to achieve greater physical fidelity through scale-bridging over brute-force increases in the number of mesh elements. For instance, quantitative predictions of transport in nanoporous media, critical to hydrocarbon extraction from tight shale formations, are impossible without accounting for molecular-level interactions. Similarly, inertial confinement fusion simulations rely on numerical diffusion to simulate molecular effects such as non-local transport and mixing without truly accounting for molecular interactions. With these two disparate applications in mind, we develop a novel capability which uses an active learning approach to optimize the use of local fine-scale simulations for informing coarse-scale hydrodynamics. Our approach addresses three challenges: forecasting continuum coarse-scale trajectory to speculatively execute new fine-scale molecular dynamics calculations, dynamically updating coarse-scale from fine-scale calculations, and quantifying uncertainty in neural network models.

The permeability of complex porous materials can be obtained via direct flow simulation, which provides the most accurate results, but is very computationally expensive. In particular, the simulation convergence time scales poorly as simulation domains become tighter or more heterogeneous. Semi-analytical models that rely on averaged structural properties (i.e. porosity and tortuosity) have been proposed, but these features only summarize the domain, resulting in limited applicability. On the other hand, data-driven machine learning approaches have shown great promise for building more general models by virtue of accounting for the spatial arrangement of the domains solid boundaries. However, prior approaches building on the Convolutional Neural Network (ConvNet) literature concerning 2D image recognition problems do not scale well to the large 3D domains required to obtain a Representative Elementary Volume (REV). As such, most prior work focused on homogeneous samples, where a small REV entails that that the global nature of fluid flow could be mostly neglected, and accordingly, the memory bottleneck of addressing 3D domains with ConvNets was side-stepped. Therefore, important geometries such as fractures and vuggy domains could not be well-modeled. In this work, we address this limitation with a general multiscale deep learning model that is able to learn from porous media simulation data. By using a coupled set of neural networks that view the domain on different scales, we enable the evaluation of large images in approximately one second on a single Graphics Processing Unit. This model architecture opens up the possibility of modeling domain sizes that would not be feasible using traditional direct simulation tools on a desktop computer.

Segmentation and analysis of individual pores and grains of mudrocks from scanning electron microscope images is non-trivial because of noise, imaging artifacts, variation in pixel grayscale values across images, and overlaps in grayscale values among different physical features such as silt grains, clay grains, and pores in an image, which make their identification difficult. Moreover, because grains and pores often have overlapping grayscale values, direct application of threshold-based segmentation techniques is not sufficient. Recent advances in the field of computer vision have made it easier and faster to segment images and identify multiple occurrences of such features in an image, provided that ground-truth data for training the algorithm is available. Here, we propose a deep learning SEM image segmentation model, MudrockNet based on Google's DeepLab-v3+ architecture implemented with the TensorFlow library. The ground-truth data was obtained from an image-processing workflow applied to scanning electron microscope images of uncemented muds from the Kumano Basin offshore Japan at depths < 1.1 km. The trained deep learning model obtained a pixel-accuracy about 90%, and predictions for the test data obtained a mean intersection over union (IoU) of 0.6591 for silt grains and 0.6642 for pores. We also compared our model with the random forest classifier using trainable Weka segmentation in ImageJ, and it was observed that MudrockNet gave better predictions for both silt grains and pores. The size, concentration, and spatial arrangement of the silt and clay grains can affect the petrophysical properties of a mudrock, and an automated method to accurately identify the different grains and pores in mudrocks can help improve reservoir and seal characterization for petroleum exploration and anthropogenic waste sequestration., Comment: 24 pages, 8 figures, submitted to Computers and Geosciences

Predicting the spatial configuration of gas molecules in nanopores of shale formations is crucial for fluid flow forecasting and hydrocarbon reserves estimation. The key challenge in these tight formations is that the majority of the pore sizes are less than 50 nm. At this scale, the fluid properties are affected by nanoconfinement effects due to the increased fluid-solid interactions. For instance, gas adsorption to the pore walls could account for up to 85% of the total hydrocarbon volume in a tight reservoir. Although there are analytical solutions that describe this phenomenon for simple geometries, they are not suitable for describing realistic pores, where surface roughness and geometric anisotropy play important roles. To describe these, molecular dynamics (MD) simulations are used since they consider fluid-solid and fluid-fluid interactions at the molecular level. However, MD simulations are computationally expensive, and are not able to simulate scales larger than a few connected nanopores. We present a method for building and training physics-based deep learning surrogate models to carry out fast and accurate predictions of molecular configurations of gas inside nanopores. Since training deep learning models requires extensive databases that are computationally expensive to create, we employ active learning (AL). AL reduces the overhead of creating comprehensive sets of high-fidelity data by determining where the model uncertainty is greatest, and running simulations on the fly to minimize it. The proposed workflow enables nanoconfinement effects to be rigorously considered at the mesoscale where complex connected sets of nanopores control key applications such as hydrocarbon recovery and CO2 sequestration., Comment: Full paper