(3a) Robust Decision Support Methodologies, Based on Smart Machine Learning, for Healthcare, Energy Systems and Food Processing

In the last few decades, several industrial sectors have started applying risk-based metrics to critical aspects of product manufacturing and optimization techniques to the solution of problems of industrial importance. This paradigm shift in industrial practice is expected to accelerate over the coming years. The growing demand for reliable and efficient strategies for optimal process management will call for further developments in data analytics, quantitative risk analysis, optimization and related methodologies. We may eventually reach a point when these techniques become an essential component of our society.

In view of these considerations, my passion for applied mathematics and statistics, and my rigorous chemical engineering education, during my MSc, PhD and postdoc years, I have developed competencies in process modeling, control and optimization, applied statistics and data analytics. More specifically, my research expertise and current research interests span the following topics:

• Dynamic modeling of chemical and (bio-)pharmaceutical processes;
• Quantification of parametric and structural uncertainty in nonlinear models;
• Risk assessment and risk-based decision making in process operations;
• Deterministic and stochastic dynamic optimization;
• Operational planning and scheduling of batch processes;
• Supply-chain and enterprise-wide optimization.

Specific examples of my research achievements include a framework for simultaneous optimization of the supply-chain and production systems of industrial gas producers [1] as well as two new strategies for deterministic/stochastic dynamic optimization of batch systems ([2] and [3]) and a novel approach to the integration of scheduling, dynamic optimization and control in multi-stage batch operations [4]. The latter are very flexible, and can accommodate virtually any conventional and unconventional batch system (e.g., reactors/bio-reactors, crystallizers, drying/freeze-drying units, evaporators/concentrators, animal test subjects and human patients) and any desired performance metric (e.g., loss/economic functions, indicators of process environmental impact and quantitative measures of therapeutic and side effects). Thus, they constitute a general, robust platform for the optimization of batch systems under uncertainty.

The research carried out over the past three years led to an innovative method for rapid estimation of the probability distribution (PDF) of the parameters of nonlinear dynamic models [5] and a strategy for dynamic selection of the most appropriate uncertainty set in stochastic dynamic optimization problems [6], which together allow real-time optimization of systems with stochastic and time-varying features (e.g., crystallizers/evaporators subject to fouling, reactors affected by catalyst deactivation, animal test subjects and human patients). These two approaches seek to exploit the rapid growth in the availability of experimental data, which makes it possible to model the underlying uncertainties that arise in a variety of real-world applications. Therefore, they constitute the backbone of a structured approach to the analysis and utilization of big data sets.

My current research efforts are devoted to developing innovative algorithms, including new optimization-driven Monte Carlo methods for PDF estimation [5], general strategies for quantification of structural model uncertainty and Bayesian approaches to online process verification and sensor health monitoring in pharmaceutical processes [7], and to the application of dynamic modelling, uncertainty quantification, and deterministic/stochastic dynamic optimization to ultrafiltration systems for concentration of monoclonal antibody solutions. The development of these new algorithms was motivated by the need for superior efficacy, reliability and computational efficiency, which are essential to achieve the principal goals of industry 4.0.

The methods and tools, developed during my MSc, PhD and postdoc years, combined with my expertise in process systems engineering, form the essence of a structured approach to the optimal utilization of experimental data for modeling, condition monitoring and optimization of both conventional and unconventional systems under uncertainty. This approach encompasses three separate steps:

• Development of hybrid mechanistic/statistical models of the systems of interest (conventional first-principles models, based on mass/energy/momentum conservation, augmented with Bayesian hierarchical components), which retain all the benefits of first-principles models, learn from experimental data as data-driven models, and provide point estimates of the variables of interest as well as estimates of their degree of uncertainty.
• Use of experimental data to train these smart machine learning models, which implies solution of an appropriate PDF estimation problem.
• Combination of trained smart machine learning models and appropriate numerical algorithms for solving condition monitoring and optimization problems under uncertainty.

In the future, I will continue augmenting this general approach for modeling, condition monitoring and optimization under uncertainty, and will apply it to the solution of challenging interdisciplinary problems in healthcare, energy systems and food processing. The latter include:

• Development of a general optimization-driven platform for optimal drug administration for human and/or veterinary applications. Potential applications of this platform include an improved artificial pancreas featuring fully automated model tuning, more reliable insulin administration policies and automatic compensation for long-term variations in the patient’s response to treatment, pediatric drug administration with automatic minimization of the risk for side effects and ideal administration of broad-spectrum antibiotics to animals.
• Development of general optimization-driven methodologies for the rational design of solid and liquid pharmaceutical formulations. These methodologies will insure optimal formulation of injectable biologics, powder preparations for production of tablets, and other innovative drug products.
• Definition of a general optimization-driven framework for rational design and optimal online management of food processing plants, comprised of components for optimal process synthesis, dynamic optimization and real-time risk-based decision making (e.g., real-time quality assurance and predictive maintenance). The potential applications of this framework encompass the production of nutrition/dietary supplements (e.g., artificial meat), the fermentation of wine/liquors, the spray drying of milk/natural extracts, the lyophilization of fruits/vegetables, etc.
• Development of a general strategy for accurate analysis of the microbial growth dynamics in packaged foods, reliable estimation of minimum/expected/maximum shelf-life, and rapid screening of different packaging material candidates. The systematic use of this strategy will benefit most entities of the food supply chain and reduce food waste.
• Definition of short-term and long-term strategies for minimizing the frequency of power outages and reducing the probability of structural damage to the power grid, caused by electric vehicles. The first consist of new grid management policies, such as the introduction of new electricity pricing schemes and the definition of different management strategies for energy storage and power generation systems, while the second represent actual modifications to the grid infrastructure, such as the installation of new production capacity and energy storage systems, combined with tailored improvements to the local distribution network.

This research portfolio is subject to modification based on funding availability, potential collaborations with other faculty members and emergence of new research challenges and/or new problems of industrial/societal relevance. An example of relevant new problem is the analysis of the temporal evolution of the COVID-19 pandemic via epidemiological models, followed by the identification of those corrective actions, which offer a satisfactory reduction in the average infection rate with the minimum economic and societal impact.

Teaching Interests:

In my MSc, PhD and postdoc years, I have developed expertise in control and optimization theory, numerical methods, applied statistics and data analytics, dynamic/steady-state process modelling and coding/parallel computing (especially in C++ and MATLAB). Therefore, I would be best prepared to teaching classes in the following areas:

• Steady-state/dynamic process modeling;
• Process design;
• Process control and optimization
• Applied statistics;
• Data analytics;
• Numerical analysis.

However, I could and would be willing to teach any class on fundamental aspects of chemical engineering (e.g., transport phenomena, thermodynamics and chemical reaction engineering). In the future, I would also like to introduce and teach two new advanced courses, whose topics lie at the intersection of chemical engineering, statistics and computer science:

• Applied statistics and big data analytics for chemical engineers;
• Applications of parallel computing to chemical engineering problems.

Since optimization and condition monitoring strategies have become essential tools for the solution of many problems of industrial relevance in recent years, I strongly believe these two interdisciplinary courses would provide chemical engineers with knowledge useful and valuable to their careers.

References:

1. Rossi F, Manenti F, Reklaitis G. A general modular framework for the integrated optimal management of an industrial gases supply-chain and its production systems. Computers and Chemical Engineering. 2015;82:84-104.
2. Rossi F, Manenti F, Buzzi-Ferraris G. A novel all-in-one real-time optimization and optimal control method for batch systems: Algorithm description, implementation issues, and comparison with the existing methodologies. Industrial and Engineering Chemistry Research. 2014;53:15639-15655.
3. Rossi F, Reklaitis G, Manenti F, Buzzi-Ferraris G. Multi-scenario robust online optimization and control of fed-batch systems via dynamic model-based scenario selection. AIChE Journal. 2016;62:3264-3284.
4. Rossi F, Casas-Orozco D, Reklaitis G, Manenti F, Buzzi-Ferraris G. A computational framework for integrating campaign scheduling, dynamic optimization and optimal control in multi-unit batch processes. Computers and Chemical Engineering. 2017;107:184-220.
5. Rossi F, Mockus L, Reklaitis G. Rigorous Bayesian inference VS new approximate strategies for estimation of the probability distribution of the parameters of DAE models. Computer Aided Chemical Engineering. 2019;46:931-936.
6. Rossi F, Manenti F, Buzzi-Ferraris G, Reklaitis G. Stochastic NMPC/DRTO of batch operations: batch-to-batch dynamic identification of the optimal description of model uncertainty. Computers and Chemical Engineering. 2019;122:395-414.
7. Rossi F, Mockus L, Manenti F, Reklaitis G. Present and future of model uncertainty quantification in process systems engineering. Chemical Engineering Transactions. 2019;74:625-630.