Digital Twin and Machine-Learning for Optimized Robotic Production of Complex Multiphase Foods - The UC Berkeley Center for Next Generation Food Systems

BACKGROUND: In the world of systems engineering, increasingly sophisticated and integrated approaches for AI systems are appearing at a rapid rate. Food production systems have however lagged in the adoption of such technology. One key component to developing a proper AI paradigm is a digital twin of a physical system.

Today, there is no shortage of general simulation software; however, the fundamental limitations are ease-of-use and integration of models with in-field data and easy deployment, tailored for food systems. A core issue in production lines is the ability of a system to adapt to rapid changes in the environment and system capabilities by autonomously modifying operation-with humans in the loop. By developing these tools for food systems researchers, we believe that we can enhance the ability of researchers to conduct research that invariably involve digital twins, as well as benefiting the nation and the world at large.

Currently, a modest level of modern technologies has been implemented in food production. For example, sensors, cameras, telecommunications have not been widely deployed. Furthermore, the cost of specialized equipment has been prohibitive and the development of coherent, easy-to-use, rapid data fusion/management systems across different platforms is lacking. Additionally, while control systems exist, they simply are too slow to be useful in deployed mobile computing platforms in harsh environments. The long-term mission of this research is to integrate and implement convergent research in the development of smart, robust, and inexpensive systems that are easy to maintain, upgrade, and deploy, incorporating state-of-the-art technologies.

SPECIFIC AIMS: The objectives are to develop Digital Twin and Machine Learning algorithms for optimized components associated with robotic production and handling of complex multiphase foods.

Project Part I

Example of precise robotic deposition of heated complex media. (Zohdi, https://msol.berkeley.edu/)

PROJECT PART I: This entails expertise of faculty in Optimization, Machine-learning, High-performance computing, Robotics, Fluid Mechanics, Micromechanics. Specifically:

(1) Use of robotic food handlers: Rapid deposition/handling of materials is an integral component of food production. This is a time-consuming method that can be streamlined with digital twin technologies. There are many components that comprise such a system, thus motivating the construction of a simulation tool framework assembling submodels of the

(a) Kinematics of the robot arm and the dispenser-printer head,

(b) Gravitational forces on the released material

(d) Conductive, convective and radiative cooling of the media

The modular system structure allows for easy replacement of submodels within the overall framework. Numerical simulations are undertaken to illustrate the overall system model, which is comprised of an assembly of submodels. Afterwards, a Machine Learning Algorithm (MLA) is developed to identify and optimize the proper system parameters which deliver a desired printed pattern.

GENERAL GOALS – PROJECT PART I: Specifically, we will develop an MLA to ascertain the appropriate robotic motion needed to create/assemble food products which would be difficult or impossible by guesswork. These models will then be used for designing high-performance feedback control systems; especially their use for predictive control and receding horizon estimation.

Accurate control of robotic motion requires accurate estimation of the interior state of the robotic manipulators in the face of poor sensor data, including noise, biases, etc. Additionally, optimal and safe decision making requires the prediction of how uncertainty will propagate in response to control actions, and what control actions will minimize cost while achieving satisfactory performance. We will combine, specifically, model-based approaches to uncertainty quantification (such as the Kalman-filter or particle filter) using large datasets. The trade-offs between the approaches, their robustness to process changes, and their computational costs will be investigated.

Project Part II

Thermal behavior of multiphase foods. (Kim, Zohdi, Singh, https://msol.berkeley.edu/)

PROJECT PART II: Across many modern industries, as technologies have matured, the use of more complex processes involving multiphase materials has increased. In the food industry, multiphase fluids are now relatively wide-spread, in particular, because of the desire to have faster throughput for large-scale food production. In many cases involving transport, such materials consist of a fluidized binder material with embedded particles. As one increases the volume fraction of particles, a corresponding increase in effective overall viscosity occurs. Often, during the process, the material must be heated, for example, to ensure food safety, induce pasteurization, sterilization, etc.

GENERAL GOALS PROJECT PART II: For real-time control, this requires rapidly computable models to guide thermal processing, for example by applied electrical induction. In the present analysis, models are developed for the required heating field (typically electrically induced) and pressure gradient needed in a delivery system (pipe-flow) to heat a multiphase material to a target temperature and to transport the material with a prescribed flow.