(417c) Design Centering through Derivative-Free Optimization
AIChE Annual Meeting
2021
2021 Annual Meeting
Computing and Systems Technology Division
Design and Operations Under Uncertainty
Wednesday, November 10, 2021 - 8:38am to 8:57am
Design space definition is one of the key components in pharmaceutical research and development. Flexibility index [3] and design centering [4] are two ways to estimate a candidate design space. Flexibility index is used to describe an operational range, which represents a maximum scaled departure of all process parameters from the given nominal conditions, such as a largest rectangle inscribed in the feasible space of steady-state operation in which control or recourse variables are adjusted. Since the design space is limited by the qualification ranges for process parameters, the result of flexibility index can be used to approximate the design space as an inscribed largest feasible region, which may be a rectangle, ellipse or other appropriate and acceptable shapes. If the nominal conditions of the process parameters are unknown, the flexibility index problem can be extended to a design centering problem, which focuses on determining the optimal nominal conditions, while maximizing the feasible operating region.
The goal of a design centering problem is to find an optimal nominal point, which corresponds to the largest flexibility index; thus, the design centering problem can be solved based on the flexibility index model. From a mathematical viewpoint, the design centering problem corresponds to a bi-level optimization model [5]. In the upper-level problem, the nominal point is searched within the feasible region; in the lower-level problem, an exact flexibility index should be calculated for each nominal point. Flexibility index problems are commonly formulated as a mixed-integer linear or nonlinear programming (MILP/MINLP) model [6-8]. For a general bi-level optimization problem, the most important issue is how to guarantee finding the global optimal solution of the lower-level model at each iteration. Similarly, for the design centering problem, the key issue is to guarantee locating the global optimal flexibility index for each nominal point.
In this work, since the design centering is a bi-level optimization problem, we first propose a novel formulation of the flexibility index for the lower-level optimization. A search direction is used to define the candidate flexibility index along a direction from the given nominal point for a specified shape of the feasible region. By applying the Karush-Kuhn-Tucker (KKT) conditions and complementarity conditions, the flexibility index model can be transformed into a single-level MINLP model. Thus, a global optimization solver (e.g., BARON) can be used to solve this model, and the exact flexibility index for a given nominal point can be obtained. This new formulation of flexibility index mainly has three features: (1) Compared with the traditional tri-level model, the proposed formulation is a bi-level model, and the corresponding MINLP model is simpler because it only requires once the KKT reformulation; (2) The traditional Haar condition [6] is not required, because the proposed model has ability to find the direction corresponding to the active constraint directly; (3) The proposed method can be extended to any shape of feasible region, as long as the representation of the shape can be provided.
For the upper-level optimization of the design centering model, a derivative-free optimization (DFO) method [9] is applied to search the nominal point in the entire feasible region. The main steps of the proposed design centering method are: (1) Perform the Latin Hypercube Sampling (LHS) strategy over the space of process parameters, where upper and lower bounds are required; (2) Check the feasibility of each sampling point. (3) Solve the DFO model of design centering problem for each feasible LHS point. The proposed flexibility index model is used to calculate the exact flexibility index at each iteration, and the obtained flexibility index for each candidate nominal point is stored. (4) The optimal nominal point is the one that has the largest value of flexibility index. In the case studies, Pyomo [10] is applied to define the models. BARON is called to solve the MINLP model through the interface of Pyomo and GAMS. Rectangular and elliptical sets for the process parameters are considered in each case. Several cases with convex and nonconvex feasible regions demonstrate the computational efficiency of the proposed design centering method.
Keywords: design space; flexibility index; design centering; derivative-free optimization.
Reference:
[1] Kusumo KP, Gomoescu L, Paulen R, GarciÌa-MunÌoz S, Pantelides CC, Shah N, Chachuat B. Bayesian Approach to Probabilistic Design Space Characterization: A Nested Sampling Strategy. Ind Eng Chem Res. 2019;59(6):2396-2408.
[2] Hakemeyer C, McKnight N, John RS, Meier S, Trexler-Schmidt M, Kelley B, Zettl F, Puskeiler R, Kleinjans A, Lim F, Wurth, C. Process characterization and design space definition. Biologicals. 2016;44(5):306-318.
[3] Swaney RE, Grossmann IE. An index for operational flexibility in chemical process design. Part I: Formulation and theory. AIChE J. 1985;31(4):621-630.
[4] Harwood SM, Barton PI. How to solve a design centering problem. Math Methods Oper Res. 2017;86(1):215-254.
[5] Colson B, Marcotte P, Savard G. An overview of bilevel optimization. Ann Oper Res. 2007;153(1):235-256.
[6] Grossmann IE, Floudas CA. Active constraint strategy for flexibility analysis in chemical processes. Comput Chem Eng. 1987;11(6):675-693.
[7] Ochoa MP, Grossmann IE. Novel MINLP formulations for flexibility analysis for measured and unmeasured uncertain parameters. Comput Chem Eng. 2020;135:106727.
[8] Pulsipher JL, Zavala VM. A mixed-integer conic programming formulation for computing the flexibility index under multivariate gaussian uncertainty. Comput Chem Eng. 2018;119: 302-308.
[9] Rios LM, Sahinidis NV. Derivative-free optimization: a review of algorithms and comparison of software implementations. J Glob Optim. 2013;56(3):1247-1293.
[10] Hart WE, Laird CD, Watson JP, Woodruff DL, Hackebeil GA, Nicholson BL, Siirola JD. Pyomo-optimization modeling in Python. Second Edition. Springer, 2017.