Checkpoints and experimental data for: Language Model Enabled Structure Prediction from Infrared Spectra of Mixtures

This repository contains pretrained model checkpoints and experimental data accompanying the paper Language Model Enabled Structure Prediction from Infrared Spectra of Mixtures. Model Checkpoints Three pretrained model checkpoints are provided. binary_align_mlp_mse_5 Model checkpoint and tokenizer. The model was trained with an MLP alignment network and MSE loss (λ = 5). This model was finetuned to produce the results in Table 1 multitask Model checkpoint and tokenizer for a multitask model trained on mixtures with component fractions ranging from 10–100%. After finetuning results of this pretrained model are reported in Table 2. ternary Model checkpoint and tokenizer for the ternary model. Results are in Table 3 Experimental Data The repository also includes the experimental data used to produce the results in the paper. Raw IR spectra are provided as CSV files in the mixture_ir_spectra and pure_ir_spectra folders for mixture and pure component spectra, respectively. Mass spectrometry data for the mixture components is provided in ms.mgf. The processed dataset used to obtain the results in Table 4 is provided in experimental_data.parquet. Finally, reference.csv contains reference information for all mixtures and compounds used in this study.