Learning from fluorescence

Abstract

We propose a systematic approach for monitoring important productivity parameters in a Dunaliella salina culture using 2D fluorescence data. For this purpose, a methodology based on Machine Learning algorithm Projection to Latent Structures Regression (PLSR) coupled with variable selection strategies was used. Additionally, a robustness analysis is proposed to support the validation of the yielded models and provide a measure of their reliability. Robust (i.e., Q2 ≥ 0.5) and parsimonious (i.e., selecting down to 3 % of the fluorescence variables present in a 250–700 nm wavelength excitation-emission matrix) models were obtained for monitoring cell count, chlorophyll b, total carotenoids and β-carotene culture concentration, and the ratio between total carotenoids and total chlorophylls, all of which were validated with a left-out batch performing with R2 higher than 0.7 except for β-carotene (R2 = 0.54).