• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Physics-Informed Neural Network Predicts Underground Hydrogen Storage Interfacial Tension Under Compositional Shifts

by
November 21, 2026
in Technology
Reading Time: 5 mins read
0
Physics-Informed Neural Network Predicts Underground Hydrogen Storage Interfacial Tension Under Compositional Shifts

Physics-Informed Neural Network Predicts Underground Hydrogen Storage Interfacial Tension Under Compositional Shifts

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A hybrid modeling framework that embeds thermodynamic principles directly into a neural network architecture can accurately predict interfacial tension across complex subsurface gas mixtures while maintaining predictive stability during compositional distribution shifts, according to a computational investigation published in Results in Engineering. Researchers Shadfar Davoodi and David A. Wood conducted a comparative analysis of four modeling strategies using an aggregated experimental database of 2,649 laboratory measurements compiled from twelve published investigations. The primary objective was addressing the fluid dynamics governing underground hydrogen storage, where injected hydrogen commingles with cushion gases such as carbon dioxide, methane, and nitrogen in porous geological formations like deep saline aquifers and depleted hydrocarbon reservoirs. Interfacial tension between the multicomponent gas phase, formation brine, and rock minerals determines capillary pressure, which governs fluid distribution, capillary trapping efficiency, and structural containment security beneath impermeable caprock formations.

Understanding interfacial dynamics under subsurface conditions remains a primary technical challenge for large-scale energy storage operations designed to balance seasonal and diurnal renewable power fluctuations. Traditional laboratory measurements of interfacial tension are expensive, time-consuming, and difficult to execute systematically under extreme reservoir conditions. Conversely, empirical correlations derived from a limited set of variables often lack mathematical flexibility, failing to capture complex non-linear behaviors across broad operating windows. Purely data-driven machine learning models frequently exhibit brittle out-of-distribution behavior, generating physically implausible predictions or violating thermodynamic laws when applied beyond their immediate training data. In real geological storage settings, injected pure hydrogen mixes with native reservoir fluids, cushion gases, and secondary gases generated by biochemical and geochemical reactions, yielding multi-gas systems that conventional correlations fail to predict reliably.

To overcome these structural limitations, the researchers formulated a physics-informed machine learning architecture that blends an explicit analytical thermodynamic baseline equation with a deep residual neural network. Unlike conventional physics-informed neural networks that enforce governing partial differential equations as hard constraints through automatic differentiation, this framework adopts a two-component decoupled design. An interpretable analytical baseline equation derived through customized formulaic optimization provides an initial physical prediction based on temperature, pressure, salinity, and individual gas mole fractions. A fully connected residual neural network then learns only the discrepancy between the analytical physics baseline and experimental measurements, preventing the network from having to relearn well-established physical dependencies while ensuring predictions remain anchored to thermodynamic priors.

The integration of the physics equation and the residual network is governed by an adaptive, prediction-level blending strategy incorporating a sigmoid-constrained parameter denoted as b. Rather than enforcing rigid algebraic combinations or unconstrained additive corrections, the parameter b dynamically modulates the magnitude of the residual neural network’s correction. During iterative training via backpropagation, the learnable weight controlling b converged consistently to a mean value of 0.623 with a standard deviation of 0.037. This mathematical result indicates that the residual neural network applies a damping factor of approximately 37.7 percent to its predicted residual correction, establishing that the mechanistic baseline provides the dominant structural foundation for the final interfacial tension value while the neural network supplies targeted non-linear adjustments.

The experimental database curated for model development encompasses 2,649 interfacial tension data points spanning temperatures from 271.2 to 448.9 Kelvin, pressures up to 45.2 megapascals (452 bar), salinities from 0 to 4.95 mol/kg, and interfacial tension values ranging between 26.6 and 85.9 mN/m. Quaternary gas mixtures consisting of hydrogen, carbon dioxide, methane, and nitrogen account for 84.5 percent of the aggregated database, reflecting representative subsurface gas mixtures, with the remainder comprising binary systems and pure hydrogen. Statistical analysis of the database revealed that gas composition exerts the strongest influence on interfacial tension, with carbon dioxide demonstrating a strong negative correlation coefficient of minus 0.62, while temperature and pressure exhibit moderate negative correlations of minus 0.32 and minus 0.37, respectively, and formation brine salinity displays a positive correlation of 0.25 driven by salting-out effects.

To guarantee strict methodological rigor and prevent data leakage, the authors implemented a systematic 60/20/20 data partitioning scheme, allocating 1,589 samples to training, 530 samples to validation, and 530 samples to primary testing. Evaluation of seven candidate split ratios demonstrated that increasing training data beyond 60 percent yielded diminishing returns in root mean square error and coefficient of determination while substantially diminishing the statistical power of the held-out validation and testing sets. Feature scaling for input parameters was conducted exclusively using RobustScaler fitted on the training split to resist outlier distortion, while target interfacial tension values were normalized using StandardScaler. All scaling parameters were subsequently applied to validation and testing partitions without re-fitting, preserving strict isolation between development and evaluation stages.

The study compared four distinct predictive paradigms: an unconstrained purely data-driven artificial neural network, an original physics-based empirical equation optimized by particle swarm optimization across ten cross-validation folds, a fine-tuned version of that empirical equation calibrated against experimental data, and the hybrid physics-informed machine learning model. During primary in-distribution testing on the held-out 530-sample test partition, the purely data-driven artificial neural network attained the highest numerical accuracy, achieving a coefficient of determination of 0.9319, a root mean square error of 3.072 mN/m, and a mean absolute error of 2.020 mN/m. The hybrid physics-informed model performed closely behind, recording a coefficient of determination of 0.9118, a root mean square error of 3.496 mN/m, and a mean absolute error of 2.575 mN/m, while substantially outperforming the original physics equation (R² of 0.5428, RMSE of 7.959 mN/m) and the fine-tuned empirical correlation (R² of 0.8031, RMSE of 5.224 mN/m).

Despite the slight numerical lead of the artificial neural network on primary in-distribution data, extensive stability and sensitivity evaluations revealed significant operational liabilities in the purely data-driven approach. Across fifty independent hyperparameter tuning runs across ten data splits, the artificial neural network exhibited high sensitivity to initialization, recording an R² coefficient of variation of 0.170, a root mean square error coefficient of variation of 0.419, and a tuning success rate of zero percent when benchmarked against a validation R² threshold above 0.80. In contrast, the hybrid physics-informed model demonstrated fourfold greater stability, achieving an R² coefficient of variation of 0.042, an RMSE coefficient of variation of 0.208, a 100 percent tuning success rate across all splits, and rapid convergence in fewer than 50 training epochs compared to more than 100 epochs for the unconstrained network.

The defining performance distinction between the modeling architectures emerged when evaluating their predictive resilience against an independent blind-test dataset that represented a distinct compositional distribution shift. The authors deployed an external dataset comprising 121 experimental measurements from Janjua and colleagues involving pure hydrogen-brine systems across temperatures of 298 to 343 Kelvin, pressures of 0.1 to 20 megapascals, and salinities between 0 and 3.0 mol/kg. Although the thermodynamic operating parameters fell within the boundary ranges of the primary database, the complete absence of cushion gases represented an acute compositional departure from the quaternary-dominated training distribution. Under this shift, the performance of the purely data-driven artificial neural network deteriorated sharply, its coefficient of determination plummeting from 0.9319 on the primary test set to 0.3423 on the blind test, accompanied by an RMSE of 4.272 mN/m and an MAE of 3.496 mN/m.

Conversely, the hybrid physics-informed neural network demonstrated superior generalization under the same compositional transition, maintaining a robust coefficient of determination of 0.7529, a root mean square error of 2.618 mN/m, and a mean absolute error of 2.146 mN/m on the blind test set. The original empirical physics equation and its fine-tuned counterpart achieved blind-test R² scores of only 0.2851 and 0.3890, respectively. These comparative findings illustrate that while unconstrained neural networks excel at interpolating complex multi-gas interactions within their familiar training domain, their lack of physical boundaries causes severe degradation when encountering compositional shifts. By anchoring the learning trajectory with an explicit analytical physics baseline and penalizing excessive residual drift via a soft loss regularization term, the hybrid framework preserves thermodynamic consistency and delivers reliable estimates crucial for assessing underground hydrogen containment integrity.

Subject of Research: Physics-informed machine learning modeling of interfacial tension in underground hydrogen storage systems

Article Title: Physics-informed machine learning for predictive modeling of interfacial tension in hydrogen–multi-gas–brine systems: A comparative analysis with primary and blind testing

Article References: Davoodi, S., & Wood, D. A. (2026). Physics-informed machine learning for predictive modeling of interfacial tension in hydrogen–multi-gas–brine systems: A comparative analysis with primary and blind testing. Results in Engineering, 32, Article 114232. https://doi.org/10.1016/j.rineng.2026.114232

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.114232

Keywords: underground hydrogen storage, interfacial tension, physics-informed machine learning, capillary trapping, cushion gas, quaternary gas mixtures, residual neural network, particle swarm optimization, compositional distribution shift, blind testing, multiphase fluid dynamics, salting-out effect, thermodynamic consistency, capillary pressure, subsurface reservoir engineering, containment security

News Source: Blake Davidson. (November 21, 2026). Physics-Informed Neural Network Predicts Underground Hydrogen Storage Interfacial Tension Under Compositional Shifts. Scienmag.

Tags: blind testingcapillary pressurecapillary trappingcompositional distribution shiftcontainment securitycushion gasinterfacial tensionmultiphase fluid dynamicsparticle swarm optimizationPhysics-Informed Machine Learningquaternary gas mixturesresidual neural networksalting-out effectsubsurface reservoir engineeringthermodynamic consistencyunderground hydrogen storage
Share12Tweet7Share2ShareShareShare1

Related Posts

Computational Design Models Metal-Free Bloch Surface Wave Optical Fiber Sensor

Computational Design Models Metal-Free Bloch Surface Wave Optical Fiber Sensor

November 22, 2026
Open-Source Framework ESFEX Matches Benchmark Consensus Across Ten Power Grid Systems

Open-Source Framework ESFEX Matches Benchmark Consensus Across Ten Power Grid Systems

November 22, 2026

Open Source QGIS Plugin Automates Multi Method Spatial Interpolation for Agriculture

November 22, 2026

Harmony Framework Unifies Modeling and Stability Analysis for Hybrid Power Grids

November 22, 2026

POPULAR NEWS

  • Meta-Analysis Finds Maize-Legume Intercropping Improves Yield and Stability in Sub-Saharan Africa

    Meta-Analysis Finds Maize-Legume Intercropping Improves Yield and Stability in Sub-Saharan Africa

    29 shares
    Share 12 Tweet 7
  • Ribociclib shows promise for specific recurrent endometrial cancer subtypes in clinical trial

    29 shares
    Share 12 Tweet 7
  • EPHRIN-B2-Based CAR-T Cells Demonstrate In Vitro Activity Against Brain Tumors

    29 shares
    Share 12 Tweet 7
  • Machine Learning Predicts Long-Term Vestibular Schwannoma Response to Radiosurgery

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Farm Data Predicts Barn Ammonia but Fails Prospective Time-Aware Testing

Meta-Analysis Finds Maize-Legume Intercropping Improves Yield and Stability in Sub-Saharan Africa

Ribociclib shows promise for specific recurrent endometrial cancer subtypes in clinical trial

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.