Lecture
New Digital Tools for Chemists: Data Infrastructure Methods and Physically Inspired Deep Learning Models"
- at -
- ICM Saal 4b
- Type: Lecture
Lecture description
T. Bocklitz; D. Jawad; J. Ravi
Photonic data science is generating tools for chemists and can be separated into a preprocessing and data modelling phase as well contains data/model storage. In spectral measurement method, etaloning distorts spectroscopic data. To address this,
an inverse modeling framework integrating computational physics with deep learning has been developed, using a two-phase transfer learning strategy [1]. This starts with pretraining on over 30,000 simulated spectra and is followed by fine-tuning with real
data. The extensive simulated dataset improves model generalization and reduces distortions by up to 70%, enhancing spectral accuracy and interpretability. The modelling of biophotonic data, particularly Raman spectroscopy, faces challenges
of limited spectral data, leading to overfitting. This is solved [2] by generating simulated spectra with quantum chemistry methods, enabling efficient pretraining of deep learning models. These are fine-tuned on smaller datasets like bacterial spectra, proving costeffective and maintaining performance comparable to models trained from scratch. This validates synthetic data's utility for model pretraining.
Finally, the model and the data need to be stored reliable. In microscopy research, integrating vast data volumes from disparate platforms is challenging. Aligning with FAIR data principles requires tools that integrate data from various sources. LEO
(Linking Electronic Lab Notebooks with OMERO) [3] addresses this, creating links between distributed systems. Initially for ELNs and OMERO, its scalable architecture now supports broader integration like the link with databases like VibSpecDB [4], enhancing collaboration and data management in photonic data science.
Literature:
[1] R. Vulchi et al., J. Chem. Inf. Model. 2025, 65, 21, 11733–11745.
[2] J. Kamran et al. J. Chem. Inf. Model. 2025, 65, 13, 6632–6643
Photonic data science is generating tools for chemists and can be separated into a preprocessing and data modelling phase as well contains data/model storage. In spectral measurement method, etaloning distorts spectroscopic data. To address this,
an inverse modeling framework integrating computational physics with deep learning has been developed, using a two-phase transfer learning strategy [1]. This starts with pretraining on over 30,000 simulated spectra and is followed by fine-tuning with real
data. The extensive simulated dataset improves model generalization and reduces distortions by up to 70%, enhancing spectral accuracy and interpretability. The modelling of biophotonic data, particularly Raman spectroscopy, faces challenges
of limited spectral data, leading to overfitting. This is solved [2] by generating simulated spectra with quantum chemistry methods, enabling efficient pretraining of deep learning models. These are fine-tuned on smaller datasets like bacterial spectra, proving costeffective and maintaining performance comparable to models trained from scratch. This validates synthetic data's utility for model pretraining.
Finally, the model and the data need to be stored reliable. In microscopy research, integrating vast data volumes from disparate platforms is challenging. Aligning with FAIR data principles requires tools that integrate data from various sources. LEO
(Linking Electronic Lab Notebooks with OMERO) [3] addresses this, creating links between distributed systems. Initially for ELNs and OMERO, its scalable architecture now supports broader integration like the link with databases like VibSpecDB [4], enhancing collaboration and data management in photonic data science.
Literature:
[1] R. Vulchi et al., J. Chem. Inf. Model. 2025, 65, 21, 11733–11745.
[2] J. Kamran et al. J. Chem. Inf. Model. 2025, 65, 13, 6632–6643
[3] Guerrero et al. Arxivhttps://doi.org/10.48550/arXiv.2508.00654.
[4] VibSpecDB https://vibspecdb.de/
