MIT’s SpectroGen: AI Generates Cross‑Modal Spectra from One Input, 99% Correlation

Transforming Spectral Data into Mathematical Distribution Curves

To match experimental fidelity, the researchers represented each spectrum as a mathematical distribution curve, employing Gaussian, Lorentzian, and Voigt functions as physical priors. This representation captures peak positions, widths, and signal characteristics, enabling the model to learn realistic spectral features.

Physics‑Guided VAE Architecture

SpectroGen builds on a variational auto‑encoder (VAE) framework. The input distribution curve is encoded into a latent variable, constrained by the physical priors, and then decoded to reconstruct the target modality (e.g., Raman). KL‑divergence loss minimizes the distributional gap between generated and real spectra, ensuring high‑fidelity output.

Accuracy Comparable to Experimental Acquisition

Using the RRUFF database (6,066 standard mineral spectra), the team selected 319 IR–Raman pairs and 371 XRD–Raman pairs for training and testing. Evaluation metrics include SSIM, RMSE, PSNR, and correlation. For IR–Raman conversion, SpectroGen achieved SSIM = 0.96 ± 0.03, RMSE = 0.010 ± 0.006, and correlation = 0.99 ± 0.01. For XRD–Raman, SSIM rose to 0.97 ± 0.04 and PSNR reached 43 ± 4 dB.

Evaluating Spectral Information Completeness

The authors further tested classification performance on 26 mineral classes across ten repeated runs. Generated spectra yielded an average classification accuracy of 90.476% (test‑set accuracy = 50.100%), whereas experimentally collected spectra achieved 69.879% (test‑set accuracy = 61.644%). The authors attribute the lower test‑set scores to the limited dataset size but note that the generated spectra still convey essential molecular vibration information.

Role of Physical Priors

Ablation experiments showed that modeling IR spectra with an incorrect Lorentzian prior or XRD spectra with a Gaussian prior caused noticeable degradation in peak height, signal‑to‑noise ratio, and peak shape, highlighting the critical contribution of physically informed priors to model interpretability and precision.

AI‑Driven Materials Science Paradigm

The study demonstrates that AI can replace costly physical instruments for spectral analysis, opening a new paradigm where machine learning accelerates both material characterization and downstream tasks such as performance prediction and application recommendation.