Machine learning aided bio-oil production with high energy recovery and low nitrogen content from hydrothermal liquefaction of biomass with experiment verification

• Dataset about HTL of biomass for bio-oil production was systematically complied. • Multi-task ML models were successfully developed for bio-oil prediction. • ML-based reverse optimization was achieved to guide high-quality bio-oil production. • Optimal solutions from reverse optimization were experimentally verified. Hydrothermal liquefaction (HTL) of biomass with high moisture (e.g., algae, sludge, manure, and food waste) is a promising and sustainable approach to produce renewable energy (bio-oil) and protect the environment. However, the production of bio-oil with high yield and preferable properties such as low nitrogen content (N_oil) is time/labor-consuming using the traditional HTL experimental method. To this context, machine learning (ML) algorithms were employed to aid the bio-oil production with the consideration of related factors in HTL, including biochemical and elemental compositions of biomass, process parameters, and solvents. Results showed that the random forest (RF) algorithm was the best one (average R 2 = 0.80) for the multi-task prediction of bio-oil yield (Yield_oil), N_oil, and energy recovery (ER_oil), hence employed for post feature interpretation and optimization. Feature importance indicates that both Yield_oil and ER_oil follow the trend of lipid content in biomass > temperature > retention time, while N_oil follows the trend of N content in biomass > temperature > retention time. Then ML-based optimization was conducted to guide the experimental research to produce bio-oil with high yield and low N content. The HTL experiment verification based on the optimal solutions from the ML-based Particle Swarm Optimization achieved the maximum Yield_oil (54.30%) and minimal N_oil (2.60%) from model biomass (composed of protein content 28%, lipid 48%, and carbohydrates 21%) at 300 °C and 30 min. The experiment verification was successful as the results were comparable to the modeling results, with errors of less than 7% for Yield_oil and ER_oil, and 23% for N_oil, and the N_oil of the experiment N_oil (2.60%) was interestingly lower than the modeled one (3.37%). This work provides new insight and strategy to accelerate the engineered HTL for desired bio-oil production.