Research map / Reinforcement learning / Visual and representation RL
Counting atoms faster: policy-based nuclear magnetic resonance pulse sequencing for atomic abundance measurement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains PPO policies that shape low-field NMR RF pulses (chirping to maximize transverse magnetization, spoiling to relax, and a toggle agent switching between them) to speed up 13C atomic abundance measurement in simulation.
Problem
Measuring atomic abundances (e.g., carbon-13) usually needs lab conditions and expensive destructive or large-scale equipment; standard NMR pulse sequences are slow, so in situ low-cost measurement is impractical.
Approach
Learn NMR pulse-modulation policies with reinforcement learning on accessible observables (transverse magnetization) rather than a full Hamiltonian, using three cooperating agents: a chirping policy to maximize signal, a spoiling policy to rapidly return spins to equilibrium, and a toggle policy that switches control between them.
Key results
- Learned chirping policy reaches maximum magnetization in roughly 16% of the time needed for a standard 1D NMR pulse sequence, with a steeper, more monotone concentration relationship.
- Learned spoiling policy achieves up to 28% reduction in average magnetization over the evaluation period versus passive relaxation.
- Toggle model outperforms chirp-only model by allowing repeated partial reset via spoiling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a learned chirp policy instead of the standard 1D NMR pulse sequence | NMR pulse sequencing · Caffeine concentration measurement | measurement time: roughly 16% of the time required | Helped |
| Pretrain the chirp policy on the standard 1D NMR pulse sequence (ablation) | Chirp policy · Caffeine samples at higher concentrations | maximum magnetization | Helped |
| Increase instrument noise progressively during training (ablation) | Chirp policy · Caffeine sample magnetization | maximum magnetization | No clear effect |
| Use progressive noise increases during training (ablation) | Chirp policy · Caffeine sample magnetization | final performance | No clear effect |
| Actively spoil transverse magnetization rather than applying no pulse | NMR relaxation between measurements · Transverse magnetization reduction | average magnetization reduction: up to 28% reduction | Helped |
| Toggle between chirping and spoiling policies | NMR pulse sequencing · Repeated magnetization measurement | maximum magnetization | Helped |
| Reset the tracked maximum magnetization to zero at toggle transitions without using the spoiling policy (ablation) | Chirp-only policy · Repeated magnetization measurement | maximum magnetization | Hurt |
Surprising findings
- Increasing training noise progressively did not meaningfully change final chirping policy performance.
- Behavior-cloned initialization from the 1D NMR pulse converged within 300K timesteps, suggesting the standard pulse may be a local minimum.
Limitations
- Evaluated only in simulation on caffeine-in-water; no real low-cost hardware experiments.
- Assumes sample nuclear distribution matches training data; requires periodic cross-checks against traditional NMR.
- Inter-nuclear couplings (e.g., proton coupling) may cause interference not modeled.
Experimental setup
Datasets: in-house caffeine-in-water NMR FIDs (6 concentrations, JEOL 500 MHz spectrometer)
Benchmarks: simulated caffeine-in-water NMR environment
Metrics: maximum magnetization vs molarity average magnetization time to max magnetization relative to 1D NMR pulse
Baselines: standard 1D NMR pulse sequence passive relaxation (no pulse) GRAPE (discussed, not compared)
Builds on: GRAPE Proximal Policy Optimization magnetic resonance fingerprinting
Closest papers by meaning
- Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra (NeurIPS 2025 · similarity 0.80)
- Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.78)
- MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.78)
- ChemPile: A 250 GB Diverse and Curated Dataset for Chemical Foundation Models (NeurIPS 2025 · similarity 0.77)
- MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.77)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.