Graphormer-IR: Graph Transformers Predict Experimental IR Spectra Using Highly Specialized Attention
Bibliographic record
Abstract
Infrared (IR) spectroscopy is an important analytical tool in various chemical and forensic domains and a great deal of effort has gone into developing in silico methods for predicting experimental spectra. A key challenge in this regard is generating highly accurate spectra quickly to enable real-time feedback between computation and experiment. Here, we employ Graphormer, a graph neural network (GNN) transformer, to predict IR spectra using only simplified molecular-input line-entry system (SMILES) strings. Our data set includes 53,528 high-quality spectra, measured in five different experimental media (i.e., phases), for molecules containing the elements H, C, N, O, F, Si, S, P, Cl, Br, and I. When using only atomic numbers for node encodings, Graphormer-IR achieved a mean test spectral information similarity ( SIS μ ) value of 0.8449 ± 0.0012 ( n = 5), which surpasses that the current state-of-the-art model Chemprop-IR ( SIS μ = 0.8409 ± 0.0014, n = 5) with only 36% of the encoded information. Augmenting node embeddings with additional node-level descriptors in learned embeddings generated through a multilayer perceptron improves scores to SIS μ = 0.8523 ± 0.0006, a total improvement of 19.7σ ( t = 19). These improved scores show how Graphormer-IR excels in capturing long-range interactions like hydrogen bonding, anharmonic peak positions in experimental spectra, and stretching frequencies of uncommon functional groups. Scaling our architecture to 210 attention heads demonstrates specialist-like behavior for distinct IR frequencies that improves model performance. Our model utilizes novel architectures, including a global node for phase encoding, learned node feature embeddings, and a one-dimensional (1D) smoothing convolutional neural network (CNN). Graphormer-IR’s innovations underscore its value over traditional message-passing neural networks (MPNNs) due to its expressive embeddings and ability to capture long-range intramolecular relationships.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".