materials-chemistry/ai produced the result/Nature Communications 2020 · v2
Neural networks read infrared spectra to work out how CO sits on platinum
Researchers trained ensembles of neural networks on simulated infrared spectra so that a measured spectrum of carbon monoxide on platinum could be turned into distributions of binding sites and local coordination.
spectrum · one line per step, placed by what the step does · bright lines used AI
Infrared spectroscopy data- and physics-driven machine learning for characterizing surface microstructure of complex materials
Nature Communications, 2020
doi:10.1038/s41467-020-15340-7 · record aix-00027 v2 · checked 2026-10-07
- AI was for
- Structure determination, Simulation surrogate
- Model family
- Multilayer perceptron, Clustering, Linear model
- Checked by
- Experimental4 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Infrared spectroscopy is a standard way to probe a surface. Molecules stuck to a surface vibrate, and those vibrations absorb infrared light at particular frequencies, producing peaks in a spectrum. The frequency depends on where the molecule sits: whether it is bonded to one metal atom, bridging two, or nested among three, and on how crowded that atom's own neighbourhood is. A real catalyst surface is not uniform. It has terraces, edges, corners and defects, so a measured spectrum is many overlapping signals added together. Pulling the mixture apart into a picture of the surface is the hard part, and interpretation has traditionally leaned on rules of thumb about which peak means which site.
The researchers worked with carbon monoxide on platinum, and also with nitric oxide as a second probe molecule. They used density functional theory, a quantum-mechanical method for calculating the behaviour of electrons, to compute vibration frequencies and intensities for a single adsorbed molecule on platinum clusters, nanoparticles and flat surfaces. That gave 1,090 sets of frequencies and intensities, filtered down to 878 stable cases. The aim was to use these as building blocks for reading real spectra.
Where AI came in
Two ensembles of 200 neural networks each were trained from scratch to take a spectrum, represented as 500 points between 200 and 2,200 reciprocal centimetres, and return probability distributions: one over the type of binding site, one over groups of generalised coordination number, a measure of how well surrounded a surface atom is. Training data was synthetic. Single-molecule spectra were added together and blurred into realistic peak shapes, each mixture paired with the site distributions that produced it. Because the 200 networks differed in their data and settings, the spread of their answers supplied the uncertainty on each prediction.
Machine learning also did two supporting jobs. Clustering sorted the near-continuous coordination numbers into discrete groups so they could be written as a distribution. Fitted scaling relations stretched the single-molecule results to crowded surfaces, standing in for quantum calculations at high coverage that the authors describe as too costly to run. The trained ensembles were then applied to four spectra digitised from earlier published experiments on carbon monoxide on platinum, and the predicted distributions compared with site assignments reported for those same samples. The reported surface structure comes from the networks; no non-machine-learning route to it is given.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Density functional theory calculations of a single CO molecule adsorbed on Pt clusters, nanoparticles and extended surfaces produced a primary dataset of 1090 sets of vibrational frequencies and intensities, which was filtered to 878 local minima. Fitted coverage scaling relations, spectral mixing and Fourier convolution expanded this into hundreds of thousands of synthetic complex infrared spectra, each paired with binding-type and generalized coordination number probability distributions. Neural network ensembles of 200 networks were trained on these synthetic spectra to predict the two distributions from a spectrum, with the ensemble spread giving prediction intervals. The trained models were applied to four digitized literature experimental spectra of CO on platinum, and the approach was also carried out with NO as a probe molecule.
How AI was used
Machine learning entered at three points in a mostly first-principles workflow. K-means clustering discretized the nearly continuous generalized coordination number into groups so that site structure could be represented as a probability distribution alongside the discrete binding-type classes. Regressed coverage scaling factors, fitted by ordinary least squares to frequencies and by nonlinear regression to intensities using extended-surface DFT data, rescaled the low-coverage DFT frequencies and intensities to arbitrary coverage, standing in for DFT calculations at high coverage. The rescaled single-adsorbate spectra were summed and convolved with randomly drawn Gaussian-Lorentzian line widths to generate the synthetic complex spectra used as training data, each discretized into 500 points at roughly 4 cm-1 intervals between 200 and 2200 cm-1. Two separate ensembles of 200 neural networks each performed multinomial regression from a spectrum to a binding-type pdf and to a GCN-group pdf, using a softmax output with a Wasserstein loss for which the authors derived a closed-form derivative; ensemble members differed in data partition, Gaussian perturbations of the primary DFT data and hyperparameters, so that the spread across members provided uncertainty estimates. The trained ensembles were then run on digitized experimental spectra to produce site-type and coordination distributions.
The shape of the work
Structural · the record, drawn
no AI
Compute DFT frequencies and intensities for single adsorbed CO
Numerical or physics simulation, including where a learned surrogate replaces it.
Our CO primary DFT dataset consists of 1090 unique sets of frequencies and intensities.where the paper describes this · verbatim
no AI
Remove unphysical outliers from the primary dataset
Cleaning, filtering, normalising or labelling data already obtained.
removing samples that are not local minima on the potential energy surface, such as transition states and saddle pointswhere the paper describes this · verbatim
AI
Label sites by binding-type and cluster GCN values into groups
Cleaning, filtering, normalising or labelling data already obtained.
K-means clustering is an unsupervised learning method that assigns GCN values to discrete GCN groupswhere the paper describes this · verbatim
AI
Apply frequency and coverage scaling factors to low-coverage data
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
we multiply frequencies and intensities calculated from DFT at low coverage (single CO) by a CSFwhere the paper describes this · verbatim
no AI
Mix and convolve spectra to build the synthetic secondary dataset
Producing candidate objects that did not previously exist.
the spectra are added together linearly (a process known as spectral mixing) to generate complex synthetic spectrawhere the paper describes this · verbatim
AI
Train neural network ensembles for binding-type and GCN pdfs
Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.
ensembles of 200 neural networks are trained on synthetic spectra generated from primary DFT datawhere the paper describes this · verbatim
AI
Infer microstructure from experimental spectra
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
After training on synthetic complex spectra, the model is applied to experimental spectra.where the paper describes this · verbatim
no AI
Compare predictions against known surface structures
Testing outputs against ground truth.
we test the structural surrogate model with coverage effects on well-defined literature experimental HREELS and surface Raman datawhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The reported surface microstructure (binding-type and GCN distributions inferred from experimental spectra) is produced by the trained neural network ensembles; no non-ML route to that result is reported.
we implement multinomial regression via neural network ensembles to learn probability distributions functions (pdfs) that describe adsorption siteswhere the paper describes this · verbatim
We display four digitized literature experimental spectra in Fig. 6awhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of Binding-type neural network ensemble (200 networks, softmax output with Wasserstein loss)Which version of the model was used is not stated.
- Version of GCN-group neural network ensemble (200 networks, softmax output with Wasserstein loss)Which version of the model was used is not stated.
- Version of K-means clustering of GCN values into discrete groupsWhich version of the model was used is not stated.
- Version of Coverage scaling factor regressions (OLS for frequencies, nonlinear fit for intensities)Which version of the model was used is not stated.
- What step 3 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00027, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error