Controlled Formation of Nanostructures with Desired Geometries. Part

Aug 8, 2015 - Sivaraman Ramaswamy, Paul I. Barton, and George Stephanopoulos*. Department of Chemical Engineering, Massachusetts Institute of ...
0 downloads 0 Views 1MB Size
Page 1 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

Controlled Formation of Nanostructures with Desired Geometries: Part 4. Multi-Resolution Optimal Control in Dynamically Directed Self-Assembly of Nanoparticles Sivaraman Ramaswamy, Paul I. Barton, and George Stephanopoulos∗ Department of Chemical Engineering, Massachusetts Institute of Technology E-mail: [email protected]

Abstract This paper presents an optimal control strategy that can guide any initial random configuration of nanoparticles to a final structure of desired geometry, in minimum time. It employs a multi-resolution view of the dynamically evolving configurations of nanoparticles, which are described through the Adaptive Finite State Projection (AFSP) approach introduced in Part 3 of this series. External charges, attracting or repelling the nanoparticles, are the controls, whose location and intensity are determined by the optimality conditions of the optimal control strategy. To ensure analytic consistency of the parametric sensitivities, during the computation of the optimal controls, and thus guarantee the optimality of the resulting control policy, a priori determination of enlarged constant projection spaces is shown to be essential. A case study illustrates the use of the proposed optimal control strategy and illuminates several of its features, such as: superiority over a static optimal solution; evasion of kinetic ∗

To whom correspondence should be addressed

1 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

traps; and effective handling of combinatorial complications arising for systems with large-size domains and many particles.

Introduction In Part 3 1 of this series, the dynamic modeling and simulation of directed self-assembly of nanoparticles was addressed. External point charges were used to move these nanoparticles and guide the self-assembly process. The dynamics of guided self-assembly were modeled with master equations, describing the time-evolution of the probability of each configuration of nanoparticles. An adaptive version of the finite state projection approach 2–4 was developed, as an adaptive model reduction technique, and it was shown to be very effective in handling the combinatorial explosion in the number of configurations, as the size of the problem increases, by retaining a small number of configurations with significant levels of probabilistic presence. Event-detection logic was used to determine the points in time when the projection space required modification (adaptation). The approach was shown to be especially effective in handling dynamic self-assembly systems with varying stiffness, since the time scales for the adaptation of the projection space perfectly conform to the varying time scale of the dynamic evolution of the system. In Part 3 we also studied the dynamic modeling and simulation of guided self-assembly under the influence of a control strategy that was based on a multi-resolution view of the evolving system. The multi-resolution view modeled the various configurations by allocating the nanoparticles into spatial regions of varying resolution. For example, initially the configurations were modeled by allocating the particles in the entire domain of the 2-d space, and then progressively to the two halves of the domain, the four quadrants of the domain and so on. At each phase the position and intensity of the external controls formed potential wells that guided the movement of the nanoparticles to the desired spatial sub-regions of the 2-d domain. 1 As the resolution of the spatial sub-regions and consequently of the descrip-

2 ACS Paragon Plus Environment

Page 2 of 34

Page 3 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

tion of the configurations increased, the repositioning of external controls with the potential wells they created guided the nanoparticles to finer and finer resolution structures, until the desired one was achieved. Furthermore, at every phase of the multi-resolution approach, the time domain was divided into multiple steps, where each step represented a different arrangement of the external control charges. The time duration and the strengths of the external charges for each time step were chosen in such a way that the probability of the system reaching that multi-resolution structure approached a pre-specified high value. The control strategy employed in Part 3 was an ad hoc strategy, possessing no specific attributes. In this paper, the control strategy will be designed to possess certain optimality properties. Although the approach is applicable for a variety of optimization objectives, in this paper we will focus the attention to minimum time optimal control strategies, i.e. control strategies which guide the self-assembly of nanoparticles from any initial random (and unknown) configuration to a configuration of desired geometry in minimum time. The mathematically rigorous optimal control strategy will determine the positioning and strength of the external charges at each time step. Note that the proposed optimal control strategy in this paper is based on an open-loop dynamic optimization, without any feedback corrections. While the use of feedback corrections within an optimal control formulation can be envisioned for future developments, there are no measurement technologies, presently available, which could monitor effectively the position of particles at the nanoscale for real-world applications. As in the companion paper, Part 3 of this series, the model for the dynamically directed self-assembly is based on an Ising lattice model. The probability of observing the nanoparticles in a particular configuration as a function of time is given by a master equation: 1,5 X X dpα (t) = − β←α (t)pα (t) + α←β (t)pβ (t), dt β6=α β6=α

∀t ∈ (t0 , tf ],

pα (t0 ) = pα,0 ,

(1)

where pα is probability of observing the system in configuration α and β←α is a propensity factor (i.e. rate) for transition from configuration α to β. Equation (1) can be written in

3 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 34

matrix form as: dp (t) = A(t)p(t), dt

∀t ∈ (t0 , tf ],

p(t0 ) = p0 ,

(2)

where p is a vector of probabilities for all considered configurations and A(t) is the corresponding matrix of propensity factors. The rates (propensity factors) β←α are calculated using the energies of the corresponding configurations, α and β. The energy of a particular configuration is computed by taking into account the pair-wise interactions among the particles and the interactions among the particles and the external charges. For example, the potential energy of a configuration, α, is given by:   V X V V N EP X X X kC (t)qp2 a(t) qs (t)qp zα,i zα,i + − zα,j , E0,α (t) = kC (t) 6 |r | |r | |r | i,s i,j i,j i=1 j6=i i=1 s=1

(3)

where V is number of cells in the domain, qs is the strength of external charge s, qp is the strength of the charge of each nanoparticle, NEP is the number of external charges, a is the parameter that lumps the effects of the close-range interactions, kC is the parameter associated with the electrostatic interactions, zα,i is a binary variable taking values 0 or 1, denoting the absence or presence of a particle in cell i of the configuration α, |ri,j | is the distance between nanoparticle i and j, and |ri,k | is the distance from the center of cell i to point charge k. Although the summation is over the cells in the domain, the inclusion of binary zα,i variables counts only those cells that have a particle in it. Note that the propensity factors will change if the external charges change. The model and the computational code for its simulation are described in more detail in. 5

Formulation of the Optimal Control Problem Selecting a control strategy that leads the nanoparticles from any initial distribution to the structure of desired geometry in minimum time, is an objective that conforms to the needs

4 ACS Paragon Plus Environment

Page 5 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

of efficient manufacturing. In addition, it ensures that the nanoparticles will avoid kinetic traps, where they are expected to stay for a long time, before they can gradually escape from the unwanted local minimum potential well and continue on to the globally minimum one of the desired structure. In addition, there are also technological limitations that constrain the placement of external charges on a physical domain. These limitations are an artifact of the various top-down approaches used for the creation of these external charges. Thus, the continuous-time optimal control problem is formulated, at every phase of the multi-resolution strategy, as follows:

min

q,r,tf

s.t.

tf dp (t) = A(q(t), r(t))p(t), dt

∀t ∈ (0, tf ],

p(0) = p0 ,

pd (tf ) ≥ pmin = pss − δ, qm ≤ q(t) ≤ qM , r(t) ∈ R, q(tf ) = qf ;

∀t ∈ [0, tf ],

∀t ∈ [0, tf ], r(tf ) = rf .

(4)

q is the vector of controls representing the ratio of the strengths of the external charges to the strength of the charge of a nanoparticle, qs /qp . r is the vector containing the locations of all controls, i.e. of the external charges. qm and qM are the lower and upper bounds, respectively, on the controls. R is a discrete set of locations on the domain where the external charges can be placed. As mentioned earlier, a discretized spatial domain is used. qf is the vector of final ratio of the strengths of the external charges to the strength of the charge of a nanoparticle, and rf is the vector of final locations of the external charges. The second constraint provides the termination criterion for the optimal control policy, where pd is the probability of the desired configuration of the nanoparticles. The value of pmin is determined by using the locations and strengths of the external charges determined by the static solution found by Solis et al., 6 which are captured by the last two equality constraints. 5 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 34

The static solution produced by Solis et al., 6 ensures robustness of the desired configuration at the steady-state, i.e. it ensures a high probability that the steady-state configuration remains at the structure with the desired geometry for a long time. As a consequence, Solis et al. showed that the static solution produces a very narrow probability distribution for the configurations, and this steady-state probability, say pss , can be easily estimated from samples generated by Metropolis Monte Carlo simulations. Fixing the locations and strengths of the charges at the final time to the static solution values ensures that the probability of the system being in the desired configuration approaches pss with time, and thus ensures feasibility of the optimal control policy. This is analogous to the use of steadystate constraints on the state (or, output), during the computation of the open-loop optimal control over the prediction horizon, within the Model Predictive Control framework, in order to ensure stability of the feedback control law. 7 For numerical solution, the controls can then be discretized in time, keeping the locations and strengths of the external charges constant within a time interval. Note, however, that the state variables, i.e. the probabilities, remain continuous in time. Each of these discretized time intervals corresponds to a step of the control strategy, as described in Part 3. 1 The new formulation is as follows:

min

qk ,rk ,tk

s.t.

tN dp (t) = A(qk , rk )p(t), dt

∀tk−1 < t ≤ tk ,

∀k = 1, 2, ..., N,

t0 = 0,

p(0) = p0 , tk−1 ≤ tk ,

∀k = 1, 2, ..., N,

pd (tN ) ≥ pmin , qm ≤ qk ≤ qM , rk ∈ R, qN = qf ;

∀k = 1, 2, ..., N − 1,

∀k = 1, 2, ..., N − 1, rN = rf ,

(5)

6 ACS Paragon Plus Environment

Page 7 of 34

where N is the number of steps in the control strategy. The solution of the above control problem would produce the optimum dynamic profile for the external charges for the guided (controlled) self-assembly process. However, it should be noted that, as the size of the physical 2-d domain increases, the number of potential locations for the external charges and the number of configurations increases combinatorially, and the solution of problem (5) becomes intractable. The Multi-Resolution Decomposition of the configuration space allows us to produce tractable computational solutions to problem (5), as will be shown in the next paragraph.

Multi-Resolution Decomposition of the Configuration Space Projection space at t3 Potential Energy

Potential Energy

Projection space at t2

A B

A B

0.02

Configurations Error

Configurations

Projection space at t1

t3

0.005

Projection space at t4

t4

t2 t1

104

Time

106

A B

Potential Energy

0.001

Potential Energy

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

A

B

Configurations

Configurations

Figure 1: Figure illustrating the Adaptive Finite State Projection method. As the simulation proceeds from t1 to t2 , the upper bound on the error due to model reduction (blue curve) increases and hits the maximum allowable value (top dashed red line), thus the projection space is expanded. The projection space expands further at t3 . Finally, when the error hits the minimum allowable value at t4 (bottom dashed red line), the projection space shrinks and causes a jump in the error due to loss of configurations with non-zero probability.

7 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

In Part 3 of this series, the Adaptive Finite State Projection (AFSP) method was developed in order to provide an effective dynamic model-reduction technique for handling the large sets of differential equations describing the probabilities of all configurations in an evolving nanostructure. The basic features of AFSP are illustrated in Figure 1. The simulation starts with an initial subset of configurations and calculates an upper bound on the error P due to model reduction with the selected subset of configurations, ε(t) = 1 − α∈A(t) pα (t), where α is a configuration in the selected subset A(t) (i.e. the projection space). This error exists because the selected configurations compose only part of the configuration space. The dynamic simulation continues till the error hits pre-specified maximum or minimum allowed values, at which point the projection space is either expanded (if the error hits the maximum allowed value) or shrunk (if the error hits the minimum allowed value). This logic ensures that the upper bound on the error always remains within the allowed range. Setting the maximum and minimum allowable values to the upper bound of the simulation error, as linear functions of time, is a design decision that is related to (a) the acceptable maximum of the upper bound in simulation error, beyond which an expansion of the set of configurations is required, and (b) the acceptable minimum of the upper bound in simulation error, beyond which a reduction of the set of configurations is desirable, in order to reduce the computational load. Consequently, setting the specific linear functions of the allowable maximum and minimum values of the upper bound in simulation error is an application-dependent decision. However, it should be noted that the AFSP approach could fail in maintaining a manageable small number of necessary configurations in the projection space. For example, inappropriately positioned controls could lead to a potential energy landscape that is relatively flat with a very large number of local minima distributed throughout the 2-d domain. In such case, the system is essentially ergodic over the whole space of configurations and we need to include in the projection space a very large number of configurations in order to model the self-assembly system with high accuracy. To overcome this difficulty, in Part 3 we

8 ACS Paragon Plus Environment

Page 8 of 34

Page 9 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

Figure 2: Figure illustrating the time-evolving potential energy landscape, as the external charges locations and their intensities change in time, leading to the movement of nanoparticles. The red circles with ‘-’ symbol indicate repulsive external charges, and the green circles with ‘+’ symbol indicate attractive external charges. All the external charges are located at the grid vertices. introduced a multi-resolution decomposition of the configuration space (Figure 2). Initially (Step-0), by using a large attractive charge in the center of the domain (Figure 2(a)), the particles are initially clustered around this charge, and the projection space contains only configurations with particles within a small circle (attraction basin) around the center. Subsequently, in Step-1, the charge in the center is mapped into two charges on either side of the center charge, and the intensities of the three charges change in time in such a way that the central well’s depth is progressively reduced, while the depth of the wells around the two new charges grows progressively larger (Figure 2(b) through 2(d)). Thus, particles initially clustered around the center progressively move horizontally towards the wells formed by the new charges. As the two new charges move away from the center, so do the clusters of particles around them. The net result is that the domain has been effectively decomposed into two disjoint sub-domains, since the probability of particles on the left side of the physical domain moving to the right side of the domain is low, and vice versa. An energy barrier has been

9 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

raised between the left- and the right-hand sides of the domain. When the movement of the charges stops and the intensities of the external charges take on their optimal steady-state values, the resulting nanostructures are stable and robust (see Solis et al. 6 ). In Phase-1, the physical domain was decomposed into two halves with the corresponding desired numbers of particles. In Phase-2 each half is further decomposed into two halves, thus further decomposing the corresponding configuration spaces (see Part 3 of this series 1 ). In early phases of the multi-resolution strategy, such as Phase-1 and Phase-2, the clustering of particles around deep potential wells effectively reduces the number of configurations that need to be included in the projection space and AFSP works very well in taming the number of configurations that need to be considered during the simulation of the self-assembling system. In later phases, the size of the independent physical domains and the number of particles in them has been significantly reduced, so that AFSP works very effectively. Actually, these reductions may allow complete enumeration of all configurations in the master equation, as it was demonstrated in an example in Part 3 of this series. 1 This multistep process constrains the movement of external charges in a way such that the basins of attractions in the energy landscape are well-defined. The optimal control problem is now formulated for every phase of the multi-resolution approach, and the positions of the external charges are no longer optimization variables in the control problem (5), thus making it more tractable. The resulting formulation of the optimal control problem for an arbitrary

10 ACS Paragon Plus Environment

Page 10 of 34

Page 11 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

phase j is given by (6):

min

qk ,tk

s.t.

tN dp (t) = A(qk )p(t), dt

∀tk−1 < t ≤ tk ,

∀k = 1, 2, ..., N,

t0 = 0,

p(0) = p0 , tk−1 ≤ tk , ∀k = 1, 2, ..., N, X pi (tN ) ≥ pmin , i∈Ωj

qm ≤ qk ≤ qM ,

∀k = 1, 2, ..., N − 1,

qN = qf .

(6)

Here Ωj is the set of goal configurations for the nanoparticles in multi-resolution phase j, among all possible configurations of the domain under consideration in phase j. The value of pmin is determined by implementation of the static solution at the final step of the multiresolution phase, which is determined using Solis et al.’s work. 6 This solution is calculated for an arbitrarily chosen configuration, among the set of goal configurations for the multiresolution phase j. As described in Part 3 of this series, 1 every phase of the multi-resolution approach starts with the placement of a strong attractive charge at the center of the domain under consideration and repulsive charges at the boundaries of this domain. This procedure attracts the nanoparticles towards the center of the domain. The time is initialized to zero at the end of this procedure, and the initial probability vector p0 is calculated by Metropolis Monte Carlo simulations. It will be later shown in the case study that the time duration of this initial step is much less than the total time of self-assembly in that phase. This optimal control formulation in (6) is modified into the following equivalent formu-

11 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 34

lation (7):

min

qk ,wk ,tN

s.t.

tN dp (t) = A(qk )p(t), dt

∀tN wk−1 < t ≤ tN wk ,

∀k = 1, 2, ..., N, w0 = 0, wN = 1,

p(0) = p0 , wk−1 ≤ wk , wk ∈ [0, 1], X pi (tN ) ≥ pmin ,

∀k = 1, 2, ..., N,

i∈Ωj

qm ≤ qk ≤ qM ,

∀k = 1, 2, ..., N − 1,

qN = qf ,

(7)

where wk are the ratios of the simulation times of all steps in the control strategy to the final time, i.e. wk ≡ tk /tN . Formulation (7) can be made dimensionless in time by scaling it with tN .

min

qk ,wk ,tN

s.t.

tN dp (τ ) = tN A(qk )p(τ ), dτ

∀wk−1 ≤ τ ≤ wk ,

∀k = 1, 2, ..., N, w0 = 0, wN = 1,

p(0) = p0 , wk−1 ≤ wk , wk ∈ [0, 1], X pi (1) ≥ pmin ,

∀k = 1, 2, ..., N,

i∈Ωj

qm ≤ qk ≤ qM ,

∀k = 1, 2, ..., N − 1,

qN = qf .

(8)

12 ACS Paragon Plus Environment

Page 13 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

Implementation The optimal control problem (8) can be solved using SNOPT, 8 a sequential quadratic programming solver. The SNOPT algorithm requires derivatives of the objective function and P the constraint; i∈Ωj pi (1) ≥ pmin .

Examining the differentiability of the formulation In Equation (2), under the AFSP method, the vector p is defined such that it contains the probabilities of all possible configurations, and the probability values of the configurations not included in the current projection space are set to zero. Similarly, A(t) is a full-sized matrix, where only the block diagonal submatrix corresponding to the configurations included in the current projection space is nonzero: β←α (t) 6= 0, ∀α ∈ A(t), β 6= α. However, during any simulation of the problem, only the configurations in the current projection space are simulated in order to reduce the computational time, yielding a system that varies in size during the course of a simulation. When the upper bound on the error hits the maximum or minimum allowed values, the AFSP method modifies the subset of configurations that are included in the reduced model of the system. If it hits the maximum allowed value, the subset of configurations expands with new configurations, and when it hits the minimum allowed value, it eliminates certain configurations from the projection space. In these instances, the state projection space changes, leading to a change in the structure of matrix A(t) in Equation (2). Note that the error crosses the maximum allowed value when the sum of the probabilities of the configurations included in the projection space falls below a certain level, or equivalently, when the difference of the sum from 1 exceeds a maximum value. Similarly, the error crosses the minimum allowed value when the sum of the probabilities of the configurations included in the projection space exceeds a certain level, or equivalently, when the difference of the sum from 1 falls below a minimum value. In other words, the structure of the ma-

13 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 34

trix, A(t), is effectively a function of the probabilities of the configurations in the projection space, leading to the following modification of the master equations: dp (t) = A(qk , p(t))p(t) = f (qk , p(t)). dt

(9)

To compute the optimal control strategy using SNOPT we need the sensitivities (i.e. the derivatives) of the objective function and of the state variables (i.e. the probabilities in the master equations) with respect to the control parameters. In addition, these sensitivities need to be continuous functions of the control parameters. As can be seen from Equations (14), the derivatives of the probabilities with respect to the control (optimization) parameters are calculated by differentiating the ordinary differential equations. The classical result for existence of sensitivities requires f to be continuously differentiable with respect to its arguments (see Theorem 3.1, Chapter V 9 ). However, whenever the error hits the maximum or minimum allowed values, the state projection space changes and so does the structure of the matrix, A(t), potentially leading to the state variables not being differentiable with respect to the control parameters. Specifically, while simulating the dynamic system with particular values of qk , if the error crosses the maximum allowed value, a configuration is added if the dp/dt value of that configuration is above a cutoff value and a configuration is removed if its dp/dt value is lower than another cutoff value. When the simulation is repeated with a small change in the value of qk , as the error hits its bounds, the change in the dp/dt value may not result in the addition/removal of the same configurations that were added/removed for the previous value of qk . This is due to the fact that the addition/removal of configurations depends on how far the dp/dt values are from the cutoff value. To illustrate diagrammatically this point, consider Figure 3. At the time point t = t∗ , assume that a change from qk to qk + δq does not change the way the projection space changes in time (see Figure 3(b), where sequence of projection spaces in time remains the same). However a change from qk + δq to qk + 2δq increases the value of dp/dt beyond the

14 ACS Paragon Plus Environment

Page 15 of 34

at t = t*

qk

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

qk+2δq qk+δq qk

(a) cutoff for changing the projection space

dp/dt

qk Old projection space New projection space

qk+δq

Total configuration space

qk+2δq

(b)

time

Figure 3: Figure illustrates the discontinuity in the right-hand side function of the ODE. At some time t = t∗ , (a) shows the graph of dp/dt vs qk for any configuration, and the corresponding change in the projection space that results from different values of qk is shown in (b). cutoff value (Figure 3(a)) and hence the projection space is modified in a different manner, as shown in Figure 3(b), resulting into a sequence of projection spaces that is different. The manner in which dp/dt affects the projection space is explained in Part 3 of this series. Therefore, for small changes in the value of qk , one cannot guarantee the same sequence of projection spaces during the dynamic simulation. Since the structure of A(t) changes based on the projection space, A(t) is not a continuous function. Thus, f in Equation (9) is not continuous in its arguments, potentially making the sensitivities undefined as well. Gal´an et al. 10 have dealt with sensitivity analysis in the case when the right-hand side function of a system of ordinary differential equations is discontinuous, by calculating a 15 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

jump in the sensitivities resulting from the discontinuities. However, their result is valid only when the sequence of discontinuities in the time domain is invariant with respect to the control parameters. Unfortunately, these requirements are not always satisfied by the AFSP strategy; the number and nature of discontinuities varies with qk . Furthermore, the method of Galan et al., 10 requires that the error does not hit the error bounds tangentially, but the AFSP imposes no such restrictions on the evolution of the error in time. It is clear from the above discussion, that the AFSP, as presented in Part 3 of this series, cannot be used as is for the computation of optimal control strategies, with provably correct optimality properties, and needs to be modified. The modified AFSP for the computation of optimal control strategies is based on the following requirements: 1. To ensure that f is continuously differentiable with respect to its arguments during every step of the control strategy, the projection space for every step is kept constant. The set of important configurations for the projection space is obtained by a onetime simulation of the system, using the AFSP method, for particular values for the strengths of the external charges. The potential variation in the projection space for when the strengths of the external charges are varied by the optimization procedure can be mitigated by having additional constraints on those strengths. The strengths used to determine this fixed projection space are chosen from the above additional constraints. 2. For any strengths of the external charges, the ratio of the simulation time of any step in the control strategy to the final time is kept constant. This is done for two reasons: (a) When time is non-dimensionalized in the formulation, this feature will force any particular step of the control strategy to start at the same time point in the scaled time domain for different values of qk . Hence, it ensures that the right-hand side of the differential equation becomes continuous in qk . (b) The constant projection space to be used in the optimization will remain valid. 16 ACS Paragon Plus Environment

Page 16 of 34

Page 17 of 34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

This is because the optimization would then adjust the values of qk such that the simulation reaches a similar physical state at the end of any particular step of the control strategy across different iterations. In other words, wk is selected a priori and kept constant in the optimization problem. The new optimization problem formulation incorporating the above features is given by:

min

qk ,tN

s.t.

tN dp (τ ) = tN A(qk )p(τ ), dτ

∀wk−1 < τ ≤ wk ,

∀k = 1, 2, ..., N, w0 = 0, wN = 1,

p(0) = p0 , X pi (1) ≥ pmin , i∈Ωj

qk min ≤ qk ≤ qk max ,

∀k = 1, 2, ..., N − 1;

qN = qf .

(10)

Additional constraints are added to the problem formulation in order to help keep the a priori determined projection space valid. Note that since the projection space is determined a priori for each step, A(qk ) no longer depends on p(t). The validity of the projection space can be easily tested a posteriori, by calculating the upper bound on the error resulting from using the projection space. This error should be less than some desired threshold. The right-hand side of the differential equations becomes: dp (τ ) = tN A(qk )p(τ ) = f (tN , qk , p(τ )), dτ

∀τ ∈ (wk−1 , wk ].

(11)

In this new formulation, the dynamic simulation during any particular step in the control strategy starts at the same time point for all values of qk . In addition, the structure of A(t) remains constant and its entries are continuously differentiable with respect to qk . Hence, f is continuously differentiable with respect to its arguments.

17 ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 34

Furthermore, in the new formulation the only potential discontinuities occur at the transition times between the various steps in the control strategy, i.e. at τ = wk . Moreover, the number of discontinuities remains fixed and they always occur in the same order. At any of the transition times, all new configurations added to the projection space start with probability values of zero, which are same as their values prior to being added to the projection space. Therefore, for all the new configurations, and the configurations that continue in the projection space, their probabilities, and hence their sensitivities, remain continuous at the transition times. This is because the time points of the discontinuities are independent of the optimization parameters, as pointed out by Gal´an et al. 10 However, the probabilities of the configurations that are removed from the projection space at the transition times jump to zero, thus causing their sensitivities to jump to zero as well, as can be shown by applying the results of Gal´an et al. 10 For the sake of simplicity, it can be assumed that the transitions between various steps in the control strategy occur at equal intervals in the scaled time domain. Hence, the final formulation becomes:

min

qk ,tN

s.t.

tN dp (τ ) = tN A(qk )p(τ ), dt



k−1 k