(229b) Deep Learning-Enabled Modeling Framework to Evaluate the Financial Feasibility of Hydrothermal Liquefaction Waste Valorization Systems

Organically rich solid waste valorization processes have the potential to zero-waste disposal and minimize greenhouse gas emission by eliminating landfill of waste as well as creating high-value-added products from it. Hydrothermal liquefaction (HTL) is a feasible process that converts biomass into an energy-dense biocrude oil that can be upgraded and refined to make liquid fuels, platform chemicals, as well as materials (e.g., bio-based asphalt binder) from various feedstocks. However, the tradeoff between deployment and scale-up options for viable feedstocks, as well as uncertainty around biocrude yield due to feedstock variability makes it difficult to determine the ideal configurations and operating parameters HTL valorization systems. To solve this problem, we incorporate a deep learning (DL) regression model that forecasts biocrude yield based on the biochemical composition of feedstock profiles. The results show the high predictability of the model with a mean squared error (MSE) < 0.02 for validation data. When integrated into techno-economic analysis (TEA), the model can be leveraged to deliver insights for cost-competitive new waste valorization pathways or upgrading of existing facilities with lower environmental impacts.

Introduction

Before scaling up or commercializing novel technologies, TEA is often carried out to evaluate the financial viability of the system and identify cost drivers for optimization. A minimum selling price (MSP) of product—the lowest price to achieve a net present value of zero—is commonly used to determine the process economic feasibility (Luo, et al., 2021) . As product yield typically has a significant impact on the process’s profitability, wet lab experiments are needed to provide accurate data for TEA. Depending on the variation of potential feedstocks and operating conditions, these experiments can be costly and time-consuming, especially when multiple system configurations exist.

With the advent of artificial intelligence (AI), data-driven models have been increasingly used in the modeling of complex reaction systems where the use of mechanistically based models is challenging (e.g., HTL valorization systems). To this end, machine learning (ML) approaches have been used to predict biocrude yield based on the characteristics of feedstocks during HTL (Cheng, et al., 2022) (Omidkar, Alagumalai, Li, & Song, 2024) (Cheng & Luo, 2022) , but these models have not been further used to provide insight on the financial feasibility of HTL systems. In this work, a regression-supervised deep learning (DL) model was developed to predict the yield of the biocrude from various HTL feedstocks, and the developed model was integrated into TEA to evaluate the economic feasibility of an HTL valorization system for the production of asphalt binder with co-production of biofuel additives and fertilizers.

DL Model Development

Literature data was leveraged to construct a preliminary dataset of biochemical composition (lipid, protein, cellulose, hemicellulose, carbohydrate, lignin, and ash) containing various feedstocks and associated HTL biocrude yield (709 data points from 190 publications). The data set used here had a small data size and uneven data distribution in a wide range (0‒100) of response variable. All data preparation and model development were performed using Python libraries, including scikit-learn for data pre-processing, dataset splitting, normalization, and missing value handling (Abraham, et al., 2014) and Tensorflow for model development (Si, Tarnoczi, Wiens, & Du, 2019) . All data were unified on a dry weight basis and normalized to a uniform scale prior to use. Unreported values were treated with a “replacement with mean” strategy. Pearson and spearman correlation coefficients between the biochemical composition features were then evaluated to eliminate features with strong linear correlations (cutoff = 0.7). KFold (value=3) was then used to split the dataset into 3 groups for cross-validation.

A sequential DL model with 4 layers was developed. Among the four dense layers, three (64 units) were equipped with ReLU activation and one (1 unit) with linear activation function. Finally, Adam Optimizer was used to optimize the model with a learning rate of 0.01 that follows stochastic gradient descent method (Kingma P. & Ba, 2015) . Several other supervised neural network algorithms, including multi-layer perceptron and support vector regression, were also evaluated, but the DL model with sequential layers offered better performance across model metrics and were used in subsequent analyses. Model output shows a generic trend to pick up rapid changes in biocrude yield due to variation in feedstock biochemical composition (Figure A). Notably, the assembled dataset had model compounds with 100% lipid, protein, or carbohydrate contents with nearly 100% biocrude yield, which was captured by the model, showing the robustness of the DL model. The model was further evaluated using a validation split of 0.33 (i.e., the validation dataset consists of 33% of the data). The general trend of both training and validation losses was similar across epochs. Although a larger difference was observed around 50 epochs (Figure B), presumably due to overfitting from the “replacement-by-mean” strategy used to offset missing value problems, both training and validation loss values remained extremely low (<0.02 across the epochs), revealing the robustness of the model in predicting biocrude yields.

TEA

The predicted yield of biocrude can then be utilized to evaluate MSP of a hydrothermal system for biobinder production (Figure C). Specifically, after gravity separation and skimming, the collected biocrude and char products were fractioned by distillation, with the light and medium fraction used as biofuel additives and the heavy fraction as biobinder. Aqueous products were treated by three potential approaches (sand filtration, electrochemical oxidation, or fungi treatment), after which the collected liquid could be used as liquid fertilizer to supplement crop nutrient (particularly N) requirements. TEA was then performed using discounted cash flow rate of return (Snowden-Swan, et al., 2017) using the predictive outputs from the DL model for feedstocks of varying biochemical compositions. Overall, this work demonstrates the utility of this integrated modeling framework to provide insights on system economic performance for feasibility analysis and process optimization.

References

Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., . . . Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn. Frontiers in Neuroinformatics, 1-10.

Cheng, F., & Luo, H. (2022). Evaluating the minimum fuel selling price of algae-derived biofuel from hydrothermal liquefaction. Bioresource Technology Reports, 100901.

Cheng, F., Belden, E. R., Li, W., Shahabuddin, M., Paffenroth, R. C., & Timko, M. T. (2022). Accuracy of predictions made by machine learned models for biocrude yields obtained from hydrothermal liquefaction of organic wastes. Chemical Engineering Journal.

Kingma P., D., & Ba, J. (2015). Adam: A Method for Stochastic Optimization. 3rd International Conference for Learning Representations. San Diego: ICLR .

Luo, H., Cheng, F., Yu, B., Hu, L., Zhang, J., Qu, X., . . . Luo, Z. (2021). Full-scale municipal sludge pyrolysis in China: Design fundamentals, environmental and economic assessments, and future perspectives. Science of The Total Environment, 795.

Omidkar, A., Alagumalai, A., Li, Z., & Song, H. (2024). Machine learning assisted techno-economic and life cycle assessment of organic solid waste upgrading under natural gas. Applied Energy.

Si, M., Tarnoczi, T. J., Wiens, B. M., & Du, K. (2019). Development of Predictive Emissions Monitoring System Using Open Source Machine Learning Library – Keras: A Case Study on a Cogeneration Unit. IEEE Access, 113463 - 113475.

Snowden-Swan, L. J., Zhu, Y., Bearden, M. D., Seiple, T. E., Jones, S. B., Schmidt, A. J., . . . Fox, S. (2017). Conceptual Biorefinery Design and Research Targeted for 2022: Hydrothermal Liquefaction Processing of Wet Waste to Fuels (No. PNNL-27186). Richland, WA (United States): Pacific Northwest National Lab. (PNNL).