London, UK
Javal Vyas
PhD Researcher - ML, Optimization, and Knowledge-Grounded LLM Systems
Machine learning | Optimization | Process control | Knowledge-grounded LLM systems
I build machine-learning, optimization, and agentic LLM systems for process control, scheduling, digital twins, and fault-tolerant decision-making. My research focuses on turning plant knowledge, validation, and simulation into reliable actions under operational constraints.
Research focus
LLM agents for control and recovery
Knowledge graphs, validators, simulation, and plant-specific constraints
Engineering thread
Optimization + ML for process systems
Scheduling, surrogate models, unit commitment, and digital twins
Public profiles
Scholar publications + open-source code
Updated research links and GitHub project work
Selected Work
Research systems and open-source projects across control, optimization, scheduling, and process intelligence.
Agentic framework that turns fault detection outputs into constraint-aware recovery plans, with simulation and deterministic validation before action.
Method
Graph RAG + DPPT
Setting
Process control
- Multi-agent monitoring, planning, action synthesis, simulation, and reprompting
- Validated recovery paths for batch and continuous process benchmarks
Semantic-AI workflow for generating operator-ready safety narratives and machine-verifiable C&E rules from process knowledge graphs.
Output
C&E logic
Lens
Semantic constraints
- Grounds LLM outputs in an ontology and controlled vocabulary
- Connects faults, symptoms, causes, and mitigation actions
Multimodal language-model workflow for extracting equipment tags and reconstructing process topology from P&ID drawings.
Venue
SCT 2026
DOI
10.69997/sct.198584
- Separates visual extraction from topology reasoning
- Targets scalable, semantically reliable P&ID digitization
Python package for solving resource-task-network scheduling problems with Pyomo, including experiment and visualization utilities.
Type
OSS package
Domain
Process scheduling
- Resource-task-network inputs
- Gantt, resource-level, and network visualizations
Optimization models and decomposition algorithms for meeting electricity demand at minimum cost under combinatorial commitment constraints.
Venue
SCT 2025
DOI
10.69997/sct.113099
- Decomposition method paired with EGRET unit commitment models
- Benchmarked across four power-system cases
OpenCV-based image analysis workflow for classifying primary crystals and agglomerates in crystallization monitoring.
- Contour features based on convexity, concavity, and circularity
- Notebook workflow for particle classification experiments
Research
Selected publications and current directions in reliable AI for process systems.
Selected publications
- Automating Cause-Effect Specification with Knowledge Graphs and Large Language Models arXiv - 2026 - preprint
- A Tutorial on Autonomous Fault-Tolerant Control Using Knowledge-Grounded LLM Agents arXiv - 2026 - preprint
- From Detection to Action: Using LLM Agents for Fault-Tolerant Control arXiv - 2026 - preprint
- From P&ID Drawings to Process Graphs: A Multimodal Language Model Approach Systems and Control Transactions - 2026 - journal article
- Optimization models and algorithms for the Unit Commitment problem Systems and Control Transactions - 2025 - journal article
- Autonomous Industrial Control Using an Agentic Framework with Large Language Models IFAC-PapersOnLine - 2025 - conference
- Leveraging LLM Agents and Digital Twins for Fault Handling in Process Plants IEEE ETFA - 2025 - conference
- Integration of Plant Scheduling Feasibility with Supply Chain Networks Under Disruptions Using Machine Learning Surrogates ESCAPE - 2024 - conference
Current directions
- Knowledge-grounded recovery agentsLLM agents that move from detection to safe action through plant context, graph retrieval, validators, and simulation.
- Semantic specifications for process safetyKnowledge-graph and LLM workflows for generating cause-effect logic, safety narratives, and machine-checkable rules.
- Process graphs from engineering artifactsMultimodal extraction of equipment, topology, and operating semantics from P&IDs to support digital twins.
Full publication list: Google Scholar .
Writing
Public writing and research notes on risk, reliability, and decision-time validation.
Writing on risk, reliability, and decision-time validation for AI systems.
A maintained list of papers, preprints, and citation metadata.
Skills
A compact stack for building reliable, testable decision systems.
Core
LLMs / Agents
Optimization
Process Systems
ML / Vision
Contact
If you are working on decision-making under uncertainty in control, ML, or process systems, I am happy to chat.
Best way to reach me: javalvyas2000@gmail.com