TLDR: A new AI method, Term Analysis with Graph Deep Q-Network (TAG-DQN), uses graph reinforcement learning to automate the complex and time-consuming process of determining atomic fine structure from spectral data. It can compute hundreds of atomic level energies in hours, achieving high accuracy (up to 95% for Co II) and significantly accelerating atomic data discovery for fields like astrophysics and fusion science.
The intricate world of atomic physics relies heavily on precise atomic data, particularly the determination of atomic fine structure from observed atomic spectra. This process, known as term analysis, is crucial for fields ranging from plasma diagnostics in industry to understanding the origins of heavy elements in astrophysics and advancing magnetic confinement fusion research. However, it’s a notoriously slow and labor-intensive task, often taking human experts years to analyze tens of thousands of spectral lines to determine thousands of fine structure level energies for a single atomic species.
A new artificial intelligence (AI) approach, called Term Analysis with Graph Deep Q-Network (TAG-DQN), is set to dramatically accelerate this bottleneck. Developed by researchers from Imperial College London, the University of Oxford, and the Russian Academy of Sciences, this method casts the complex decision-making of term analysis as a Markov decision process (MDP) and solves it using graph reinforcement learning (GRL).
The Challenge of Atomic Fine Structure Determination
Traditional term analysis involves extracting energies and total electron angular momenta of energy levels from observed atomic spectra. While spectrum measurement and theoretical calculations can be completed in weeks, the subsequent analysis often stretches into months or even years. This is because scientists must deduce level energies from an immense number of energy differences, guided by less accurate theoretical predictions. Existing tools assist with parts of the process, but the core decision-making remains a human endeavor, requiring deep expertise and extensive manual labor.
Introducing TAG-DQN: An AI Solution
TAG-DQN aims to automate this challenging task. It represents the atomic data—known and unknown energy levels (nodes) and spectral lines (edges)—as a graph. The AI agent then learns to make sequential decisions to determine unknown levels and match observed spectral lines to theoretical transitions. The learning process is guided by reward functions that are trained on historical human expert decisions, essentially teaching the AI to prioritize actions that lead to confident and consistent level determinations.
The system operates through two alternating action types: first, selecting an unknown level to determine its observed energy, and second, matching at least two spectral lines from the experimental line list to connections between the selected unknown level and already known levels. Graph Neural Networks (GNNs) are employed as a powerful learning representation to process the complex graph structures, enabling the AI to understand the relationships between levels and lines.
Also Read:
- FragmentRetro: A Faster Approach to Chemical Synthesis Planning
- AutoLit: An AI Approach to Faster Literature Reviews
Promising Results and Future Impact
The researchers evaluated TAG-DQN on existing spectral line lists and theoretical calculations for Co II and Nd II-III, which are elements of significant interest in atomic physics. The results were highly encouraging:
- Hundreds of level energies were computed within hours, a stark contrast to the months or years typically required by human analysis.
- For Co II, the determined level energies agreed with published values in 95% of cases.
- For Nd II-III, the agreement ranged from 54% to 87%.
- TAG-DQN consistently outperformed baseline search algorithms, including greedy search and Monte-Carlo tree search (MCTS), particularly in its ability to identify levels that remained consistent with observations and theory throughout the analysis.
An important finding from ablation studies was the critical role of “duelling” network architectures within TAG-DQN, which proved essential for handling the large action spaces involved in the problem.
This new AI approach represents a significant step towards closing the gap between the growing demand for atomic data from astronomy and fusion science and the current efficiency of atomic fine structure determination. While human validation and refinement of the AI’s outputs are still recommended, TAG-DQN sets the stage for a future where fundamental atomic data can be developed in years rather than decades. The methodology is also adaptable to other experimental techniques, such as grating spectroscopy, with minor adjustments.
This work highlights the immense potential of graph reinforcement learning and artificial intelligence to assist scientific discovery, particularly in data- and labor-intensive scientific tasks. For more details, you can read the full research paper here.
