lai_xu

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
lai_xu [2026/10/12 19:35] – Lai Xulai_xu [2026/10/14 08:42] (current) – Lai Xu
Line 4: Line 4:
  
 Email: [[lxu@bournemouth.ac.uk]] Email: [[lxu@bournemouth.ac.uk]]
-{{tag>"ARTIFICIAL INTELLIGENCE" "LARGE LANGUAGE MODELS" "AGENTIC AI" "LangChain" " Agentic System" "Retrieval-Augmented Generation (RAG)" "AtomAgents"}}+{{tag>"ARTIFICIAL INTELLIGENCE" "LARGE LANGUAGE MODELS" "AGENTIC AI" "LangChain" " Agentic Systems" "Retrieval-Augmented Generation (RAG)" "AtomAgents" "Python"}}
  
 Building on the EU-funded Horizon SUSHEAS project (https://susheas.eu/), I would like to supervise the following project topics: \\ Building on the EU-funded Horizon SUSHEAS project (https://susheas.eu/), I would like to supervise the following project topics: \\
-**Title 1: LLM-Powered Academic Assistant and Information Extraction System for Materials Science via Prompt Engineering and Fine-Tuning** \\+**Topic 1: LLM-Powered Academic Assistant and Information Extraction System for Materials Science via Prompt Engineering and Fine-Tuning** \\
 Develop a knowledge retrieval and structured data extraction Agent tailored for material scientists. The goal is to help researchers rapidly extract "Composition-Phase-Yield Strength" data triplets from thousands of messy PDF literature papers. Develop a knowledge retrieval and structured data extraction Agent tailored for material scientists. The goal is to help researchers rapidly extract "Composition-Phase-Yield Strength" data triplets from thousands of messy PDF literature papers.
 1. Learn LangChain/CrewAI frameworks and build a foundational Retrieval-Augmented Generation (RAG) pipeline for scientific papers. 1. Learn LangChain/CrewAI frameworks and build a foundational Retrieval-Augmented Generation (RAG) pipeline for scientific papers.
 2. Construct Agent logic using materials science open datasets to accurately extract structured text and multi-component composition tables from PDFs. 2. Construct Agent logic using materials science open datasets to accurately extract structured text and multi-component composition tables from PDFs.
 3. Develop a lightweight Web UI allowing users to query in natural language (e.g., "Find all austenitic stainless steels with yield strength > 200MPa that are cobalt-free"), which the Agent automatically visualises into a data table. 3. Develop a lightweight Web UI allowing users to query in natural language (e.g., "Find all austenitic stainless steels with yield strength > 200MPa that are cobalt-free"), which the Agent automatically visualises into a data table.
-4. Thesis writing, system integration testing, and viva/presentation preparation. <p> +4. Thesis writing, system integration testing, and viva/presentation preparation.  
-**Title 2: An Automated Grid-Screening Agent System for Compositionally Flexible Alloys (CFAs) via Rule-Based Interception** \\+**Topic 2: An Automated Grid-Screening Agent System for Compositionally Flexible Alloys (CFAs) via Rule-Based Interception** \\
 Re-implement the material design pipeline described in Section 2 of the multi-component stainless steel paper using an Agentic AI state machine framework. Since the Fe-Cr-Ni-Mn system resides in a manageable 4D space, grid-based brute-force filtering will be explored. Re-implement the material design pipeline described in Section 2 of the multi-component stainless steel paper using an Agentic AI state machine framework. Since the Fe-Cr-Ni-Mn system resides in a manageable 4D space, grid-based brute-force filtering will be explored.
 1. Master LangGraph, focusing on the concepts of state management, Nodes (actions), and Conditional Edges (routing logic). 1. Master LangGraph, focusing on the concepts of state management, Nodes (actions), and Conditional Edges (routing logic).
 2. Write Python scripts to wrap thermodynamic phase diagram calculators (using the open-source pycalphad library or creating a Mock API to simulate Thermo-Calc responses). 2. Write Python scripts to wrap thermodynamic phase diagram calculators (using the open-source pycalphad library or creating a Mock API to simulate Thermo-Calc responses).
 3. Wire nodes in LangGraph to form a pipeline: Grid Generation -> Phase Stability Filter (FCC > 99%) -> Martensite Start Temperature (Ms) Interception (If Ms >= 350K, reroute/reject) -> Solid Solution Strength Calculator (Varvenne Formula). Ensure the system processes ~27,000 candidate coordinates autonomously. 3. Wire nodes in LangGraph to form a pipeline: Grid Generation -> Phase Stability Filter (FCC > 99%) -> Martensite Start Temperature (Ms) Interception (If Ms >= 350K, reroute/reject) -> Solid Solution Strength Calculator (Varvenne Formula). Ensure the system processes ~27,000 candidate coordinates autonomously.
-4. Evaluate computational pipeline throughput and efficiency, finalize data logging, and draft the undergraduate thesis.<p>+4. Evaluate computational pipeline throughput and efficiency, finalize data logging, and draft the undergraduate thesis.
  
-**Title 3: High-Throughput Active Learning Alloy Design Agent Powered by Batch Bayesian Optimisation**\\+**Topic 3: High-Throughput Active Learning Alloy Design Agent Powered by Batch Bayesian Optimisation**\\
 Address the high-dimensional space challenge presented in advanced materials (like 17-element systems). Implement an Active Learning Agent using Batch Bayesian Optimization instead of grid searching to let the AI intelligently navigate complex compositional landscapes. \\ Address the high-dimensional space challenge presented in advanced materials (like 17-element systems). Implement an Active Learning Agent using Batch Bayesian Optimization instead of grid searching to let the AI intelligently navigate complex compositional landscapes. \\
 1. Study advanced Bayesian Optimization libraries in Python (e.g., Ax/Botorch or scikit-optimise). 1. Study advanced Bayesian Optimization libraries in Python (e.g., Ax/Botorch or scikit-optimise).
 2. Build a Batch BO Agent capable of proposing "q" distinct, promising recipes simultaneously (e.g., generating 50 "good enough" candidates per iteration instead of just one point) to allow for parallel compute scaling. 2. Build a Batch BO Agent capable of proposing "q" distinct, promising recipes simultaneously (e.g., generating 50 "good enough" candidates per iteration instead of just one point) to allow for parallel compute scaling.
 3. Implement multi-round termination criteria (e.g., maximum 3-5 iterations). Compare the efficiency and computational resource savings of the "Intelligent Navigation Agent" against traditional brute-force grid searching. 3. Implement multi-round termination criteria (e.g., maximum 3-5 iterations). Compare the efficiency and computational resource savings of the "Intelligent Navigation Agent" against traditional brute-force grid searching.
-4. Algorithm hyperparameter tuning, efficiency analysis, and thesis completion. <p> +4. Algorithm hyperparameter tuning, efficiency analysis, and thesis completion.  
-**Title 4: Semi-Automated Agentic System for Post-Experiment Data Analysis and Computational Model Self-Correction**\\+**Topic 4: Semi-Automated Agentic System for Post-Experiment Data Analysis and Computational Model Self-Correction**\\
 Re-engineer the Section 3 (Experimental Validation) and Section 4 (Discussion/Reflection) phases of the alloy literature. Build a "Virtual Data Scientist Agent" that handles raw experimental files and updates model parameters.\\ Re-engineer the Section 3 (Experimental Validation) and Section 4 (Discussion/Reflection) phases of the alloy literature. Build a "Virtual Data Scientist Agent" that handles raw experimental files and updates model parameters.\\
 1. Learn Python scientific libraries (Pandas, Matplotlib, Scipy) to process real-world materials data, such as noisy tensile stress-strain curves and XRD diffraction text streams. 1. Learn Python scientific libraries (Pandas, Matplotlib, Scipy) to process real-world materials data, such as noisy tensile stress-strain curves and XRD diffraction text streams.
 2. Program the "Analysis Agent" to ingest a messy .csv from a tensile tester, automatically calculate the \(0.2\%\) yield strength offset and ultimate tensile strength (UTS), and generate publication-quality figures. 2. Program the "Analysis Agent" to ingest a messy .csv from a tensile tester, automatically calculate the \(0.2\%\) yield strength offset and ultimate tensile strength (UTS), and generate publication-quality figures.
 3. Develop the "Reflection Agent." It compares experimental yields against the theoretical predictions made in Project 2/3. Using the reasoning capability of LLMs, it computes discrepancies and auto-generates a standardized engineering report outlining Calibration Factors for the next batch. 3. Develop the "Reflection Agent." It compares experimental yields against the theoretical predictions made in Project 2/3. Using the reasoning capability of LLMs, it computes discrepancies and auto-generates a standardized engineering report outlining Calibration Factors for the next batch.
-4. Thesis writing, case-study compilation, and defense presentation. <p> +4. Thesis writing, case-study compilation, and defense presentation.  
-**Title 5: Customization of the Open-Source MIT "AtomAgents" Framework for Multi-Component Material Simulations**\\+**Topic 5: Customization of the Open-Source MIT "AtomAgents" Framework for Multi-Component Material Simulations**\\
 Stand on the shoulders of giants by cloning, modifying, and deploying MIT’s open-source AtomAgents repository, adjusting its multi-agent framework to execute molecular dynamics simulations on unique material systems (e.g., Fe-Cr-Ni-Mn).\\ Stand on the shoulders of giants by cloning, modifying, and deploying MIT’s open-source AtomAgents repository, adjusting its multi-agent framework to execute molecular dynamics simulations on unique material systems (e.g., Fe-Cr-Ni-Mn).\\
 1. Set up the local Linux/server environment, clone AtomAgents, configure the Microsoft AutoGen backend, and successfully reproduce their baseline group-chat agent simulation. 1. Set up the local Linux/server environment, clone AtomAgents, configure the Microsoft AutoGen backend, and successfully reproduce their baseline group-chat agent simulation.
  • lai_xu.1791833710.txt.gz
  • Last modified: 2026/10/12 19:35
  • by Lai Xu