2026 Research Notes
24 research notes
Can physics-guided machine learning make wearable electrochemical biosensor calibration more reliable?
A lightweight neural network adds a monotonicity constraint to electrochemical calibration, reducing error on a public wearable sweat-biosensor dataset while preserving physically plausible response trends.
In brief: The physics-guided model reported RMSE 0.9077 versus 1.8547 for linear regression and R2 = 0.7940 versus 0.1400, corresponding to about a 49% reduction in RMSE.
Can machine learning identify how axial ligands tune single-atom catalysts for CO2-to-formate conversion?
A small DFT dataset trains ML models that screen hundreds of metal-ligand combinations, while SHAP and electronic-structure calculations explain why the best axial ligand depends on the metal center.
In brief: An FNN trained on 81 DFT structures reached R2 = 0.94 and RMSE = 0.14 V on the independent test set; DFT rechecks of the top 17 candidates averaged 0.05 V prediction error.
Can machine learning turn phase-field simulations into a usable map of lithium dendrite risk?
A phase-field model separates diffusion and electromigration effects on lithium deposition, then Gaussian-process regression maps local electrochemical conditions to Coulombic efficiency.
In brief: The simulations link higher voltage or current density with stronger dendritic branching, while coordinated increases in diffusion coefficient and ionic conductivity suppress tip growth; Gaussian-process regression maps these conditions to Coulombic efficiency.
Can catalyst screening become more realistic by learning operating conditions and structural ensembles?
A graph-neural-network workflow screens 22,599 dual-atom ORR structures, but ranks metal pairs using electrochemical surface states and the fraction of active configurations rather than a single idealized structure.
In brief: The model screened 22,599 structures, found 3,431 with predicted ORR limiting potentials above 0.8 V, and identified only eight of 729 metal pairs with more than 80% active configurations.
Can multiple LLM agents learn electrocatalyst design rules instead of only ranking candidates?
MAESTRO places specialized LLM agents inside an iterative catalyst-design loop, using a machine-learning force field for rapid evaluation and accumulated design history for in-context learning.
In brief: Across thousands of generated single-atom catalyst structures, the multiagent loop identified candidates that computationally break conventional ORR scaling relations and move below the roughly 0.36 V scaling-limited overpotential.
Can machine learning turn a small first-principles catalyst set into useful rules for CO2 electroreduction?
DFT screens 24 transition-metal single-atom catalysts on C6N6, while regression and classification models connect atomic descriptors with adsorption and product selectivity.
In brief: Among 24 metal centers, 19 activated CO2 and 10 favored CO2RR over HER; Fe@C6N6 gave the lowest reported limiting potential, about 0.39 V. The best SVR adsorption model reached R2 = 0.93 on the held-out test set.
Can a large machine-learning catalyst model make a prediction that survives an actual electrochemical experiment?
Open Catalyst Project adsorption-energy calculations screen bimetallic surfaces and point to PdAu3, which is then synthesized and tested for oxygen reduction.
In brief: Rotating-disk experiments showed an improved ORR onset potential for polycrystalline PdAu3 nanoparticles compared with commercial polycrystalline Pt/C.
Can machine learning create a useful reference-electrode signal without putting a reference electrode in the battery?
A virtual reference electrode estimates negative-electrode potential from ordinary two-electrode signals and uses the estimate to detect lithium-plating risk.
In brief: Reported errors were RMSE 0.023 V and MAE 0.018 V; adaptive charging extended cycle life 8.25× versus CCCV under the tested low N/P-ratio condition.
Can an electrochemical biosensor recognize its environment before estimating concentration?
A two-stage ML workflow first classifies pH from DPV shape, then routes the signal to a pH-specific concentration model.
In brief: The pH classifier exceeded 85% accuracy across four environments from pH 3.6 to 8.0; pH-specific regressors estimated CEA over five orders of magnitude.
Can a neural network choose an equivalent-circuit class from an impedance spectrum?
CNNs classify Nyquist plots before conventional parameter fitting, with transfer learning and an experimental pyrrole electropolymerization demonstration.
In brief: The published article reports >80% classification accuracy for spectra generated with random parameters.
Can catalyst knowledge be transferred from simple alloys to high-entropy surfaces instead of learning the hardest chemistry from scratch?
A sequential transfer-learning workflow moves from unary and binary systems to ternaries and then high-entropy alloys, with active learning used in the most data-scarce regime.
In brief: The transferred models reduced adsorption-energy prediction error relative to models trained directly on the limited HEA dataset for CO2-reduction intermediates.
Can a machine-learning surrogate make multiobjective PEM-electrolyzer design fast enough to optimize?
Surrogate models learn from a physical PEMWE simulation and optimize two-dimensional catalyst-loading distributions for current and thermal uniformity.
In brief: The optimized design increased modeled current density from 1.28 to 1.46 A cm−2, a reported 14.5% improvement, while retaining acceptable current and temperature uniformity.
Can the sensor materials be designed to make machine learning better at separating overlapping biomarkers?
Two polymer-modified electrochemical channels intentionally reshape biomarker voltammograms, giving ML more orthogonal information for multiplexed sensing.
In brief: Across 450 single- and mixed-analyte measurements, multi-output regression achieved an overall R² of about 0.95 for four biomarkers.
Can an electrochemical sensor plus machine learning turn COD into a 10-second online measurement?
A nitrogen-etched boron-doped diamond electrode is paired with machine learning for real-time chemical-oxygen-demand monitoring and industrial validation.
In brief: The reported system performs online COD detection in about 10 s with 96% accuracy and was deployed in an industrial setting.
Can machine-learning potentials make atomistic electrochemical-interface simulations large enough to become more useful?
MACE interatomic potentials trained through active learning reproduce DFT-level behavior across 21 charged metal–electrolyte systems.
In brief: The MLIP committee models aqueous NaCl at Au, Cu, and Rh(111) interfaces over a range of surface charge densities with DFT-level accuracy.
Can machine vision teach us what the voltage noise of gas-evolving electrodes means?
A lightweight CNN tracks microbubbles optically, creating a baseline that lets frequency-domain voltage analysis recover characteristic bubble residence times.
In brief: Bubble residence times obtained from optical tracking agreed with features in the electrochemical voltage signal, and the frequency method was extended to gauzes, foams, and felts at industrial current densities.
Can an algorithm decide which battery electrolyte experiments are worth doing next?
Deep active learning selects lithium-metal electrolyte experiments, then transfers the learned component relationships into larger and modified formulation spaces.
In brief: Three active-learning iterations used 128 tested cells and about 4.4% of the stated exhaustive-testing resources; selected formulations enabled roughly three-times longer Li||Li cycling than the reference state-of-the-art electrolytes.
What changes when an electrochemical experiment can choose its own next conditions?
A closed-loop flow-electrochemistry platform combines Bayesian optimization with online UHPLC-DAD analysis to optimize medicinally relevant reactions autonomously.
In brief: The platform was demonstrated on three distinct reaction campaigns, including C–N coupling, anodic amino-acid modification, and decarboxylative alkylation.
Can machine learning help identify which interfacial descriptors control CO2-reduction microenvironments?
A CuAg tandem-catalyst study combines AIMD, regression ML, and descriptor selection to connect imidazolium-salt microenvironments with C–C/C–N coupling energetics.
In brief: XGBoost and SISSO identify coupling energy and the average deviation in elemental ground-state band gaps as important descriptors for the limiting-potential target.
Can machine learning recover selectivity when the electrochemical sensor chemistry is not perfectly selective?
Four PFAS produce imperfectly selective DPV responses, yet a machine-learning classifier distinguishes them with 95% accuracy.
In brief: A classifier trained on differential-pulse-voltammetry data reached 95% accuracy for PFOS, PFOA, PFBS, and PFBA.
How much catalyst composition space can be searched from only 70 experiments?
XGBoost learns OER overpotential from 70 FeCoCrMnCu LDH compositions and points to a composition that is then synthesized and tested.
In brief: The model reached R² = 0.84 and RMSE = 9.95 mV; it predicted 261.3 mV for the selected composition versus 270 mV experimentally at 10 mA cm−2.
Can a pretrained machine-learning potential find a practical noble-metal-free OER catalyst in a large atomic configuration space?
An equivariant ML model screens 3,976 single-atom oxyhydroxide configurations and nominates W1–NiFeOOH for experimental water electrolysis.
In brief: W1–NiFeOOH delivered 13.1 A cm−2 at 2.0 V and operated for more than 500 h at 1 A cm−2 in AEM water electrolysis.
Can published PEM-electrolyzer data guide the design of lower-iridium anodes?
Literature data are converted into a supervised-learning problem, then used to nominate lower-Ir anode compositions for experimental testing.
In brief: Ir0.9M0.1O2 candidates with M = Cr, Zn, Sb, or Ho were experimentally tested and preserved cell performance while reducing Ir content.
Can deep learning separate two electrochemical signals that overlap and change with time?
A modified U-Net separates dopamine and octopamine in fast-scan cyclic voltammetry by learning the time–voltage–current structure of the signal.
In brief: NRMSE 0.06 for dopamine and 0.08 for octopamine on computational mixtures; experimental 1:1 mixtures gave r = 0.93 and CCC = 0.93.