Hydrogen with Carbon Management

AI/ML Approaches for Research and Development Application

Approaches
Use Machine Learning (ML) approaches based on known labels/expectations, usually for
  • Classification (e.g., Logistic, Support Vector Machine)
  • Regression (e.g., Linear, Kernel Ridge)
  • AI-assisted field site assessment for efficient hydrogen injection/withdrawal cycling under constraints
  • Operation optimization, e.g., using AI/ML to learn from operational data on H₂ hub and increase throughput, reduce emissions and costs
  • Combustion dynamics of carbon-free fuels differ from those of natural gas
  • Identifying H₂ combustion system’s dynamic stability limits (predicting flashback/blowout)
  • Predict impacts of high temperatures and pressures
  • Analyze factors in H₂ embrittlement of steels/welds
  • Leverage data to improve component design
  • Structured numeric data, text, image, maps, sequences, time-series data
  • AI/ML decision support system to assess processes and optimize operations
  • Structural and functional materials discovery and development system (e.g., for proton-conducting fuel cells and electrolyzers, catalysts for reversible solid-oxide fuel cells, oxidation-resistant alloys and coatings)
  • Real-time prediction and monitoring system for control of unit operations, structural & functional material development
  • Infrastructure planning and operation with risk assessment capabilities
Approaches
Use machine learning approaches without known labels/expectations, usually for
  • Clustering
  • Dimension reduction
  • Screen materials to narrow search by application
  • Reduce the dimensions and complexities of the data
  • Image/characterize microstructures across scales
  • Detect and group trends and anomalies
  • Detect failures in critical systems
  • Structured or unstructured data
  • Anomaly detection and early warning system
  • Data reduction modules/libraries/approaches to reduce the energy, emission, and training data to process data and train AI/ML models
Approaches
Use complex neural network architecture for
  • Image and spatial prediction
  • Time series modeling/prediction
  • Natural language processing (NLP)
Can also be used for Transfer Learning or Reinforcement Learning
  • Accelerate data processing of large sensor dataset
  • Transfer learning from subsurface models for oil and gas as well as CO₂ storage to H₂ storage models
  • Incorporate big data from operating sensors in a hydrogen hub (to predict material corrosion, infrastructure maintenance needs, etc.)
  • Extract and organize information from unstructured data
  • Predict impacts of high temperatures and pressures
  • Model material/carrier life or performance
  • Image/characterize microstructures across scales
  • Leverage Hub data to improve operations/lower
  • Optimize long-term H₂ storage as energy resource
  • Create streamlined proxy models
  • Model material degradation/microstructural changes
  • Detect/avert failures in critical systems
  • Predict degradation/changes in catalysts & cathodes
  • Structured numeric data, text, image, maps, sequences, time-series data
  • Supply chain planning and operation optimization system
  • Uncertainty and risk assessment system
  • Manufacturing process optimization system with computer vision and sensors to avoid potential microstructure and material defects
  • Maintenance and repair prediction system with knowledge to predict future incidents and potential impact
  • Service representative system equipped with NLP and domain knowledge to recommend potential solution or documentation
  • Infrastructure planning and operation with risk assessment capabilities
  • Deep learning digital twins
Approaches
Combine more than one ML models to handle the tasks (tree-based, such as Random Forest, XGBoost, or deep learning, such as Committee Machine)
  • Combustion dynamics of carbon-free fuels differ from those of natural gas
  • Identifying H₂ combustion system’s dynamic stability limits (predicting flashback/blowout)
  • Functional materials degradation root-cause analyses
  • Analyzing subsurface/near-surface geological chemistry changes after H₂ injection (subsurface storage) to indirectly detect H₂ leaks
  • Improve model performance by indicating most valuable data addition/subtraction
  • Image/characterize microstructures across scales
  • Develop low-temp., proton-conducting Fuel Cells (FCs) and electrolyzers
  • Detect/avert failures in critical systems
  • Structured numeric data, text, image, maps, sequences, time-series data
  • AI/ML decision support system to assess processes and optimize operations
  • Structural and functional materials discovery and development system (e.g., for proton-conducting fuel cells and electrolyzers, catalysts for reversible solid-oxide fuel cells, oxidation-resistant alloys and coatings)
  • Value of Information (VOI) assessment to quantify the costs (monetary or non-monetary) of adding data/information
  • Real-time prediction and monitoring system for control of unit operations, structural & functional material development
Approaches
Integrate scientific equations and knowledge into ML algorithms (usually using Deep Learning), such as Physics-Informed Neural Networks (PINNs)
  • Combustion dynamics of carbon-free fuels differ from those of natural gas
  • Identifying H₂ combustion system’s dynamic stability limits (predicting flashback/blowout)
  • Predicting impacts of high-temperature H₂ coexisting with water vapor at high pressure
  • Identifying alternative hydrogen subsurface storage options, beyond salt domes (caverns)
  • PINNs to balance physics vs ML data-driven models
  • Identify/avoid dynamic system stability limits
  • Estimate wave speed in rotating detonation combustors
  • Identify turbine stall/surge limits using physical data
  • Create streamlined proxy models
  • Explore operating extremes
  • Structured numeric data, text, image, maps, sequences, time-series data
  • Scientific equations, field equations, partial differential equations
  • Structural and functional materials discovery and development system (e.g., for proton-conducting fuel cells and electrolyzers, catalysts for reversible solid-oxide fuel cells, oxidation-resistant alloys and coatings)
  • High-fidelity AI/ML digital twins with embedded scientific knowledge for what-if scenario exploration
Large Language Models for UNSTRUCTURED DATA MINING Large Language Models for GENERATIVE DISCOVERY Transfer Learning Science-based Learning Edge Computing Advance Sensor Reinforcement Learning Operator Learning
Characterize Materials
Develop Structural/Functional Materials
Enable Smart Components w/embedded sensors
Optimize regional Planning and Operations
Modeling/Simulation Bridging Scales
Accelerate Flow Sheet Initialization
Optimize Turbine Operations
Assess/Control H2 Combustion Characteristics