arXiv:2002.11447v2 [math.OC] 8 May 2020 FP)frcytladml htfr h ocle two- called so the form that melt and crystal for (FBPs) The 2005). al., et (Jurisch Indium-Phosphide or (GaAs) h etclGain ree(G)cytlgot pro- growth crystal (VGF) Freeze Gradient Vertical The 03 04 xedtefltesbsdmto lnigto planning motion flatness-based the extend 2004) 2003, esi sdfrtepouto fhg ffiinybl com- bulk efficiency Gallium-Arsenid high like crystals of single production semiconductor the pound for used is cess h w-hs ae hl Kn n aaa,19)and 1995) Zabaras, and (Kang while case, al., two-phase et the (Rudolph problem, based full 2019) al., the et Regarding (Koga approaches. backstepping- using and or designs (Maidi 2016) feedback geometry- Corriou, 2014), by for 2012, as design al., well et feedforward (Petrus as enthalpy- the (2003), one- due al. for called et justified available Dunbar so lately often case, the are special this is yields results Regarding (OPSP). extent) (which problem spatial Stefan constant phase dominant is its phase to dis- distribution broadly one temperature is the in systems. that it parameter assumption system distributed the the of Making framework of the extent in spatial cussed the to Due is which 1984) (Crank, problems nonlinear. e.g. inherently (TPSP) boundary problem free Stefan phase of coupled Modelling two crucible. yields the system in surround power top which the the the manipulating heaters by to the done bottom is of the This from manner. grows desired such verti- a crystal plant a the single molten, through a is all moved that crucible is After the gradient chunks. in temperature seed) semiconductor cal the which solid to crucible (up with material symmetric filled rotationally placed later a is of is crystal bottom seed the A follows: at as works basically process cat(F)[rjc ubrW 4412/1-1]. WI number ⋆⋆ [project (DFG) schaft ⋆ ulcto ne raieCmosLcneCC-BY-NC-ND Licence Commons Creative a under publication hswr a enfne yteDush Forschungsgemein- Deutsche the by funded been has work This c 00teatos hswr a enacpe oIA for IFAC to accepted been has work This authors. the 2020 Keywords: Abstract: aktpig rcigcnrl ueia methods numerical control, tracking backstepping, rbesadi ia ato aycytlgot processes growth diffe crystal a addition, many In kernel of transformations growth. backstepping during part the melt of vital and approximation crystal a of is enco extent is c and problem which Stefan two-phase problems problem A Stefan process. growth two-phase crystal Freeze a of control tracking .INTRODUCTION 1. etclGain ree w-hs tfnpolm distr problem, Stefan two-phase Freeze, Gradient Vertical hscnrbto rsnsabcsepn-ae tt fee state backstepping-based a presents contribution This ∗ elhSine,MdclIfraisadTcnlg,66 Ha 6060 Technology, and Informatics Medical Sciences, Health nttt fCnrlTer,Tcnsh nvriä Dresden, Universität Technische Theory, Control of Institute ∗∗ rsa rwhpoesvabackstepping via process growth crystal oto fteVria rdetFreeze Gradient Vertical the of Control nttt fAtmto n oto niern,Uiest o University Engineering, Control and Automation of Institute tfnEcklebe Stefan io,Asra(mi:[email protected]) (email: Austria Tirol, rse,Gray(mi:{stefan.ecklebe, (email: Germany Dresden, jan.winkler}@tu-dresden.de) ∗ rn Woittennek Frank e Hnee l,20)adestepolmfo h ieof side the from problem the address 2009) al., et (Hinze h anojcieo hscnrbto st rsn a present to is contribution this of objective main The umr n notokt ute oki given. a is new Finally, work further a results. before to as simulation outlook discussed, an well presents is and summary briefly as them 6 equations solve kernel Section to existence the scheme the the computation for states 5, Section solutions and In as it of law. well for on control as tracking dynamics Based system resulting error target on. suitable later corresponding a which control the derives tracking 4 trajectories Section the and derived these, for design the Sec- control required of before feedforward are properties introduced, the some is recites states briefly process pa- 3 the distributed tion of one-dimensional detailed simplified model more a rameter 2 and publication. Section fashion forthcoming In a brief space, in rather limited given a be the will in to results done Due back- is presented. time-variant nu- this is resulting the kernels the for stepping of approach approximation different meric a state Furthermore, backstepping-based by feedback. TPSP the for control tracking Structure and Objective 1.1 different present problem. las authors flatness-based the remaining designs and the backstepping-based energy- rendering 2019) approaches, using designs al., feedback remained. et output an inputs both via (Ecklebe using tracking same TPSP In the complete the to the does of gap control 2019) a However, boundary Krstic, approach. one and energy-shaping at (Koga actuation with that noteworthy TPSP Lyapunov-based and be is the a to states for it already law has context 2010) control this FBPs al., feasi- et In two (Petrus not here. that the is account between variant into coupling one-phase taken the the since extension for ble direct approaches a the feedback, of Concerning control. optimal ∗∗ spresented. is nit ftoculdfe boundary free coupled two of onsists a Winkler Jan nee nteVria Gradient Vertical the in untered etapoc o h numerical the for approach rent btdprmtrsystems, ibuted-parameter u otetime-varying the to due bc einfrthe for design dback ∗ 01069 lin ll f t z z˜ abrupt change of the physical parameters between crystal ˜ and melt. Top Heater Γl Γl

Melt 2.2 Decomposition Assuming piecewise constant parameters for the solid Interface and the liquid phase it is possible to decompose the tempearature distribution via Crystal γ(t) 0 T (z,t), z ∈ Ω = (Γ ,γ(t)) T (z,t)= s s s (2) T (z,t), z ∈ Ω = (γ(t), Γ ) Bottom Heater  l l l ϕ with the temperatures Ts(z,t) and Tl(z,t) in the solid and Γ ˜ r s Γs liquid part, respectively. This yields the two FBPs 2 ∂tTs(z,t)= αs∂z Ts(z,t) (3a) Fig. 1. Schematics of the cylindrical coordinates (r, ϕ, z) δs ∂zTs(Γs,t)= us(t) (3b) and the moving coordinate frame z˜ = z − γ(t). λs Ts(γ(t),t)= Tm (3c) 2. MODELLING 2 ∂tTl(z,t)= αl∂z Tl(z,t) (3d) As the foundation for model based control, this section δl ∂zTl(Γl,t)= ul(t) (3e) introduces a one dimensional distributed parameter model λl of the VGF process plant. Tl(γ(t),t)= Tm (3f) where the indexes “s” and “l” denote the solid and liquid 2.1 Plant Model phase, respectively. Furthermore, the heat flows us(t) and ul(t) at the bottom and the top boundary are considered The quantity under consideration is given by the spatial as system inputs with the orientation factors δs = −1 and and temporal distribution of the system temperature T in δl = 1. For reasons of clarity, the partial derivative of the crucible, denoted in cylindrical coordinates with radius T (z,t) wrt. z and t are given by ∂zT (z,t) and ∂tT (z,t), r, angle ϕ as well as height z, and depending on the time respectively. Finally, αs = λs/(ρscp,s) and αl = λl/(ρlcp,l) t. Within this contribution, we assume that the lateral denote the thermal diffusivities. heaters are used as active isolation, avoiding any heat loss in radial direction and therefore yielding a temperature Next, examining the energy balance at the interface γ(t) distribution which is independent of r. Since the plant is yields the Stefan condition (Stefan, 1891) also rotationally symmetric, this justifies averaging over ρmLγ˙ (t)= λs∂zTs(γ(t),t) − λl∂zTl(γ(t),t) (4) the longitudinal cross-sectional area, reducing the spatial which describes the evolution of the phase boundary. domain to a line whose boundaries are represented by the Herein, ρm denotes the density of the melt at melting bottom and top of the crucible at z = Γs and z = Γl, temperature and L the specific latent heat. respectively. This yields two areas given by the crystal and the melt, separated by the moving phase boundary γ(t) Together, (3) and (4) form the TPSP whose state is given (cf. Figure 1). In contrast to the temperature distribution by in the crystal, which can be modelled via diffusion, the liq- T (·,t) x(·,t)= ∈ X = L ([Γ , Γ ]) × (Γ , Γ ) , (5) uid melt also enables convective heat transport. However, γ(t) 2 s l s l since the considered semi-conductors posses small Prandtl   numbers (e.g. 0.068 for GaAs), the dominating heat trans- where L2([Γs, Γl]) denotes the space of real-valued square- port mechanism is diffusion. Therefore, convective effects integrable functions defined on [Γs, Γl]. Note that the PDE- in the melt are neglected. ODE-PDE system defined by (3) and (4) is inherently nonlinear since the domains of (3a) and (3d) depend on Summarising, the temperature distribution in the system the state variable γ(t). is given by the distributed variable T (z,t) and governed by a one dimensional nonlinear heat equation (Cannon, 1984) Since the systems (3a)–(3c) and (3d)–(3f) share the same structure, the following sections will merely discuss generic ∂ ρ(T (z,t))cp(T (z,t))T (z,t) = variables, denoted by the ◦ symbol if the results are appli- ∂t cable to both phases. If terms from two different phases ∂ ∂  λ(T (z,t)) T (z,t) ,z ∈ (Γ , Γ ) \{γ(t)} (1) are to appear in the same expression, the complementary ∂z ∂z s l phase is marked by the • symbol.   with the density ρ, the specific heat capacity cp, and λ the thermal conductivity being temperature-dependent. Note, 2.3 Moving Coordinates that while the temperature at the interface T (γ(t),t) is fixed at the melting point temperature T due to the m To simplify the notation of the controller error system later ongoing phase transition, the heat flow in this description on, the coordinate transformation is not continuous at the phase boundary due to the release of latent heat within the solidification process and the T˜(˜z,t)= T (z,t) with z˜ := z − γ(t) (6) is introduced which maps the current interface position to holds for all n in N0. the origin of a moving frame as shown in Figure 1. This yields the generic system in the new coordinates Thus, given the definition of the flat output (12) the refer- R+ ˜ 2 ˜ ˜ ence interface trajectory γr(t) is from Gℵ≤2( ). Moreover, ∂tT◦(˜z,t)= α◦∂z˜T◦(˜z,t) +γ ˙ (t)∂z˜T◦(˜z,t) (7a) as the reference temperature distribution T˜r(˜z,t) is com- δ◦ ˜ ∂z˜T˜◦(Γ˜◦,t)= u◦(t) (7b) puted via (9) and (8), by construction Tr(˜z,t) belongs to λ◦ + Kℵ≤2(Ω˜, R ) as defined below. T˜◦(0,t)= Tm (7c) Definition 2. (Class Kℵ(Ωz, Ωt)). A function (z,t) 7→ γ˙ (t)= s◦∂z˜T˜◦(0,t)+ s•∂z˜T˜•(0,t) (7d) f(z,t) is an element of the function class Kℵ(Ωz, Ωt) if ∞ f(·,t) ∈C (Ωz) and f(z, ·) ∈ Gℵ(Ωt). where Γ˜◦ = Γ◦ − γ(t), s◦ = −δ◦λ◦/(Lρm), and s• = −δ•λ•/(Lρm). Note, that in these coordinates the interface 4. STATE FEEDBACK velocity γ˙ (t) directly enters the PDE (7a) in form of a convection coefficient. The following section states the main result of this contri- 3. FEEDFORWARD CONTROL bution concerning the backstepping-based state feedback. 4.1 Error System This section outlines a feedforward control that originates from Dunbar et al. (2003) for the OPSP and was extended For the system (7), let the error coordinates be given by by Rudolph et al. (2003) to the TPSP. Since the trajec- ˜ ˜ tories for T (z,t) and γ(t), which are computed by this e˜◦ (˜z,t)= T◦(˜z,t) − T◦,r(˜z,t) (14a) feedforward scheme, will be used as a reference in following ∆γ(t)= γ(t) − γr(t), (14b) sections, this recap is merely focussed on their properties. yielding the nonlinear error dynamics 2 ∂te˜◦ (˜z,t)= α◦∂ e˜◦ (˜z,t)+ ∆γ ˙ (t)∂z +γ ˙ (t)∂z e˜◦ (˜z,t) Since the solution T˜◦(˜z,t) of (7) can be expressed in terms z˜ ˜ r ˜ ˜ of an infinite power series in z˜, given by + ∆γ ˙ (t)∂z˜T◦,r(˜z,t )
(15) ∞ z˜i as shown in Appendix A. Linearising (15) around the T˜◦(˜z,t)= c◦ (t) , (8) ,i i! reference (e◦(˜z,t) ≡ 0, ∆γ(t) = 0) yields the linearised i=0 X error dynamics substitution into (7a) and comparison of the coefficients of 2 ∂ e˜◦ (˜z,t)= α◦∂ e˜◦ (˜z,t) +γ ˙ (t)∂ e˜◦ (˜z,t) like powers in z˜ yields the recursion formula t z˜ r z˜ ˜ 1 + ∆γ ˙ (t)∂z˜T◦,r(˜z,t) (16) c◦,i+2(t)= ∂tc◦,i(t)−γ˙ (t)c◦,i+1(t) i =0,..., ∞ . (9) as well as α◦ Examining (8) indicates that the initial
coefficients of the ∆γ ˙ (t)= s◦∂z˜e˜◦(0,t)+ s•∂z˜e˜•(0,t). (17) series are given by Hence, substituting (17) in (16) gives the linear time- variant error system c◦,0(t)= T˜◦(0,t)= Tm, c◦,1(t)= ∂z˜T˜◦(0,t) . (10) 2 Next, using (7d) as the defining equation for the melt- ∂te˜◦ (˜z,t)= α◦∂z˜e˜◦ (˜z,t) +γ ˙r(t)∂z˜e˜◦ (˜z,t) gradient + b◦(˜z,t)∂z˜e˜◦(0,t)+ c◦(˜z,t)∂z˜e˜•(0,t) (18a)

1 ∂z˜e˜◦(Γ˜◦,t)=˜u◦(t) (18b) ∂z˜T˜l(0,t)= λs∂z˜T˜s(0,t) − ρmLγ˙ (t) , (11) λl e˜◦(0,t)=0 (18c)   the solution for both phases can be expressed by the ∆γ ˙ (t)= s ∂ e˜ (0,t)+ s ∂ e˜ (0,t) (18d) ˜ ◦ z˜ ◦ • z˜ • gradient in the crystal ∂z˜Ts(0,t) and the growth rate γ˙ (t). ˜ ˜ Thus, (7) via the parametrisation (8) is differentially flat with b◦(˜z,t) = s◦∂z˜T◦,r(˜z,t), c◦(˜z,t) = s•∂z˜T◦,r(˜z,t), and with a flat output the new input u˜◦(t). Note that herein, the solid and liquid temperature errors are coupled over the whole domain by y (t) ˜s y(t)= 1 = ∂z˜T (0,t) . (12) means of their fluxes through the phase boundary. y2(t) γ(t)     Furthermore, reference trajectories for the components 4.2 Hopf-Cole Transformation yr,i(t) of yr(t) are chosen as transitions between stationary 0 e Since the coefficient γ˙ (t) of the convection term cannot states yr,i and yr,i for start and end, respectively, via r 0 e 0 be treated via the classic backstepping transform, it is yr,i(t)= yr,i + (yr,i − yr,i)Φ (t) i =1, 2 (13) eliminated by a Hopf-Cole transformation (Hopf, 1950) with Φ(t) sufficiently smooth. e˜◦ (˜z,t)= Ψ◦ (˜z,t)¯e◦(˜z,t) (19) Analysing the specific convergence conditions for (8), γ˙ r(t) and choosing Ψ◦ (˜z,t) = exp − α◦ z˜ , which is a stan- (Rudolph et al., 2003, 2004) show that it is sufficient to 2 R+ dard procedure in these cases. This yields the system demand Φ(t) ∈ Gℵ≤2( ) with the Gevrey class Gℵ(Ω) 2 from the definition below. ∂te¯◦(˜z,t)= α◦∂z˜e¯◦(˜z,t)+ r◦(˜z,t)¯e◦(˜z,t) ¯ Definition 1. (Gevrey-class; Gevrey 1918). A smooth func- + b◦(˜z,t)∂z˜e¯◦(0,t)+¯c◦(˜z,t)∂z˜e¯•(0,t) (20a) tion t 7→ f(t) defined on the open set Ω ⊂ R is an element ˜ γ˙r(t) ˜ −1 ˜ of the Gevrey class G (Ω) of order ℵ over Ω if there exists ∂z˜e¯◦(Γ◦,t)= e¯◦(Γ◦,t)+ Ψ◦ (Γ◦,t)˜u◦(t) (20b) ℵ 2α◦ a positive constant D such that e¯◦(0,t)=0 (20c) n n+1 ℵ sup|∂t f(t)|≤ D (n!) ∆γ ˙ (t)= s◦∂z˜e¯◦(0,t)+ s•∂z˜e¯•(0,t) (20d) t∈Ω where results as Section 6 shows. Certainly, for both variants 1 2 the target systems properties can only be conveyed if r0(˜z,t)= − 2¨γr(t)˜z +γ ˙r (t) (21a) 4α◦ the inverse transformation of (23) exists. This can be ¯ −1 assumed since it is of Volterra-type and therefore always b◦(˜z,t)= Ψ◦ (˜z,t) b◦(˜z,t
) (21b) −1 invertible (cf. Heuser 1992) or shown by a simple fixed- c¯◦(˜z,t)= Ψ• (˜z,t) c◦(˜z,t). (21c) point argument 1 . Note, that the resulting system now exhibits reactive terms Finally, by examining (22b), eliminating the target terms that are driven by the reference interface velocity and ac- via (23) and substituting (20b), the control input for the celeration. Furthermore, the original Neumann boundary original system with the kernel in original coordinates condition now appears as a Robin boundary condition. k◦(z,ζ,t)= k˜◦(˜z,ζ,t) is given by 4.3 Backstepping Transformation λ◦ γ˙r(t) u◦(t)= k◦(Γ◦, Γ◦,t)+ δ◦ν◦ − × δ 2α To enforce proper tracking of the reference, the errors in ◦ "  ◦  temperature and boundary position should converge to T◦(Γ◦,t) − T◦,r(Γ◦,t) + ∂zT◦,r(Γ◦,t) zero. This demand is formulated in the target system Γ◦
2 + ∂zk◦(Γ◦,ζ,t)+ δ◦νk◦(Γ◦,ζ,t) × ∂tw◦(˜z,t)= α◦∂z˜w◦(˜z,t)+ µ◦(˜z,t)w◦(˜z,t) (22a) Zγ (t) ∂z˜w◦(Γ˜◦,t)= δ◦ν◦w◦(Γ˜◦,t) (22b)
γ˙ r(t) w◦(0,t)=0 (22c) T (ζ,t) − T◦ (ζ,t) exp (ζ − Γ◦) dζ . (26) ,r 2α with the reaction coefficient µ (˜z,t) and boundary gain ν  ◦  # ◦ ◦
as design parameters. To map the system (20) into (22) the transformation z˜ 5. WELL-POSEDNESS AND NUMERICAL w◦(˜z,t)=¯e◦(˜z,t) − k˜◦(˜z,ζ,t)¯e◦(ζ,t) dζ (23) SOLUTION OF THE KERNEL EQUATIONS Z0 is used. Computing the requirements on the transforma- In this section, a numerical solution scheme for the ker- ˜ tion kernel k◦(˜z,ζ,t) (cf. Appendix B) yields the kernel nel equations (24a)–(24c) is discussed 2 . To this end, the equations integral form of the system is derived in a first step. As 2 2 stated in (Jadachowski et al., 2012), the method of succes- ∂tk˜◦(˜z,ζ,t)= α◦ ∂ k˜◦(˜z,ζ,t) − ∂ k˜◦(˜z,ζ,t) z ζ sive approximations as introduced in (Colton, 1977) and + a◦(ζ,t)k˜◦(˜z,ζ,t)  (24a) extended in (Meurer and Kugi, 2009) does not show good z˜ convergence for time-varying kernels and is therefore only 2α◦k˜◦(˜z, z,t˜ )= a◦(ζ,t) dζ − 2¯b◦(˜z,t) (24b) employed to investigate the existence of a solution. Thus, 0 the presented solution will be based on a spatial discreti- Z z˜ ˜ ¯ ˜ ¯ sation of the kernel. However, in contrast to (Jadachowski α◦k◦(˜z, 0,t)= b◦(ζ,t)k◦(˜z,ζ,t) dζ − b◦(˜z,t) (24c) et al., 2012) the Midpoint rule will be used which eventu- 0 Z z˜ ally leads to an iterative solution scheme that maintains 0= c¯◦(ζ,t)k˜◦(˜z,ζ,t) dζ − c¯◦(˜z,t) (24d) the structural properties of the problem. Z0 where a◦(˜z,t) = µ◦(˜z,t) − r◦(˜z,t). Examining the sys- tem (24), one observes that the problem for k˜◦(˜z,ζ,t), 5.1 Integral Form given by (24a)–(24c) is well-posed (cf. 5.2). Herein, the integral boundary condition (24c) arises since the term Introducing the normal form coordinates η =z ˜ + ζ and ¯ b◦(˜z,t)∂z˜e¯◦(0,t) is to be eliminated from (20a). However, σ =z ˜ − ζ yields the dynamics the demand for completely decoupled target systems and, ∂ k¯(η,σ,t)=4α∂ k¯(η,σ,t)+ a η−σ ,t k¯(η,σ,t) (27a) thus, the elimination of c¯ (˜z,t)∂ e¯ (0,t) that results in t ησ 2 ◦ z˜ • η (24d) renders the problem overdetermined. Therefore, to ¯ r ¯
4αk(η, 0,t)= a 2 ,t dr − 4b (0,t) (27b) recover a well-posed formulation the convective coupling 0 at z˜ = 0 is reintroduced with the modified target system Z η ¯ ¯ η+s
¯ ¯ dynamics αk(η,η,t)= b 2 ,t k(η,s,t) ds − b (η,t) (27c) 0 2 Z ∂tw¯◦(˜z,t)= α◦∂ w¯◦(˜z,t)
z˜ of the transformed kernel k¯(η,σ,t) = k˜(˜z,ζ,t). Hence, + µ◦(˜z,t)¯w◦(˜z,t)+ d◦(˜z,t)∂z˜w¯•(0,t) (25) formal integration of (27a) wrt. σ and η as well as z˜ where d◦(˜z,t) = c¯◦(ζ,t)k˜◦(˜z,ζ,t) dζ − c¯◦(˜z,t) replaces substitution of (27b) (derived wrt. η) and (27c) yields (24d). Obviously, by choosing µ◦(˜z,t) ≤ 0 ∀(˜z,t) and R ν ≤ 0 this approach yields exponentially stable error 1 dynamics for the one-phase case where the gradient in Note, that to rigorously proof the invertability, the transformation the adjacent phase vanishes from (17). For the two-phase has to be defined as a linear map on an appropriately chosen Banach space in order to make the underlying fixed point theory applicable. case, stability of the resulting error dynamics has to be For lack of space, details are omitted within this contribution. shown due to the bilateral coupling via d◦(˜z,t). This 2 As only variables of one phase occur in the kernel equations, the will be addressed in a forthcoming publication due to generic placeholder ◦ is dropped in favour of a more compact notation lack of space. However, simulation studies yield promising from now on. η σ σ ¯ 1 r ¯ k(η,σ,t)= a 2 ,t + ∂tk(r,s,t) 4α σ k¯ (t) Z  Z0  Nσ ,Nσ
− a r−s ,t k¯(r,s,t) ds dr 2 k¯ (t)    i,j 1 σ
+ ¯b σ+s ,t k¯(σ,s,t) ds − ¯b (σ, t) . (28) σ = η d =0 α 2 Z0 
d =1 5.2 Existence

As already presented in (Colton, 1977) a solution for (28) d =2 ∆σ ∆η by means of the method of successive approximations can k¯ (t) σ =0 k¯ (t) η be established by considering the series 0,0 Nη ,0 ∞ k¯(η,σ,t)= K¯ n(η,σ,t) (29) Fig. 2. Discretised kernel domain with selected kernel n=0 elements and required derivative orders for the com- X where the K¯ n(η,σ,t) are given by K¯ 0(η,σ,t)=0, putation of ki,j (t) (shaded). η 1 K¯ 1(η,σ,t)= a r ,t dr − ¯b (σ, t) (30a) Next, the uniform step width ∆ = ∆ = ∆ , motivated by 2 α η σ Zσ the similar dynamics in both directions, is chosen. Hence, σ
¯ n 1 ¯ σ+s ¯ n−1 approximating the integrals in (28) by lower sums yields K (η,σ,t)= b 2 ,t K (σ,s,t) ds α 0 the explicit computation scheme Z η σ 1
i−1 j−1 + ∂ k¯n−1(r,s,t) ∆2 4α t k¯ (t)= ∂ k¯ (t) − a n−m ∆,t k¯ (t) Zσ Z0  i,j 4α t n,m 2 n,m n=j m=0 r−s ¯ n−1 X X   − a ,t K (r,s,t) ds dr (30b) i−1
2 ∆ 1  + a n ∆,t − ¯b (j∆,t) Furthermore, Meurer and Kugi
(2009) show that the con- 4α 2 α n=j vergence conditions on the series (29) depend on an upper X
l ¯ n j−1 bound for ∂tK (η,σ,t) due to the repetitive differentiation ∆ + ¯b j+m ∆,t k¯ (t) (31) in (30b). Although the detailed appearance may differ from α 2 j,m the systems discussed in (Meurer and Kugi, 2009) or (Izadi m=0 X
et al., 2015), the terms that are to be examined are a(˜z,t) for the interior where 0

K article. /

2 300 9 mm L k / ) )

ke(z,t)kL2 t ( REFERENCES z,t γ ( 200 ∆γ(t) 6 e ∆ k Cannon, J. R., 1984. The One-Dimensional Heat Equation. 100 3 Vol. 23 of Encyclopedia of Mathematics and its Appli- cations. Addison-Wesley. Colton, D., 1977. The solution of initial-boundary value 0 0 0 10 20 30 40 50 problems for parabolic equations by the method of t / h integral operators. Journal of Differential Equations 26 (2), 181 – 190. Crank, J., 1984. Free and Moving Boundary Problems (Ox- Fig. 3. Trajectories of the temperature errors L2 norm (solid) and the interface deviation ∆γ(t) (dashed). ford Science Publications). Oxford Science Publications. Oxford University Press. 6. SIMULATION SETUP AND RESULTS Dunbar, W. B., Petit, N., Rouchon, P., Martin, P., 2003. Motion planning for a nonlinear Stefan problem. Eu- This section briefly presents simulation results for the two- ropean Series in Applied and Industrial Mathematics phase case. The simulated process goal is the growth of (ESAIM): Control, Optimization and Calculus of Varia- a GaAs single crystal in a 400 mm furnace. As simulation tions 9, 275–296. model, a lumped, FE-based approximation with 41 nodes Ecklebe, S., Woittennek, F., Winkler, J., Frank-Rotsch, per phase has been used. For further details on the C., Dropka, N., 2019. Towards model based control of physical parameters of the simulation setup, please refer the Vertical Gradient Freeze crystal growth process, to (Ecklebe et al., 2019, Sec. 7). available at https://arxiv.org/abs/1908.02519. Gevrey, M., 1918. Sur la nature analytique des solutions In this case, the growth process shall start at an initial des équations aux dérivées partielles. Premier mémoire. length of 200 mm and end at about 300 mm. To do so, Annales scientifiques de l’École Normale Supérieure 35, yr,2(t) is chosen as a smooth transition (cf. (13)) between 129–190. these values with the targeted solid phase gradient at −1 URL http://eudml.org/doc/81374 the interface held constant at yr,1(t) = 17Km during Heuser, H., 1992. Funktionalanalysis Theorie und Anwen- the transition (t < 25 h). Thus, the recursion (9) from dung, 3rd Edition. Teubner, Stuttgart. Section 3 yields the reference temperature trajectory t 7→ Hinze, M., Pätzold, O., Ziegenbalg, S., 2009. Solidification Tr(·,t) which in turn enables the computation of the of a GaAs melt—Optimal control of the phase interface. kernels for the solid and liquid phases by means of (34). Journal of Crystal Growth 311 (8), 2501 – 2507. Herein, 81 discrete points and the parameters µ(˜z,t) = Hopf, E., 1950. The partial differential equation ut + −1 × 10−2 s−1 as well as ν = 0 m−1 have been used in uux = µuxx. Communications on Pure and Applied the target systems of both phases (cf. (22)). Finally, the Mathematics 3 (3), 201–230. control input is computed via (26). Izadi, M., Abdollahi, J., Dubljevic, S. S., 2015. PDE For demonstration purposes, an initial error for the phase backstepping control of one-dimensional heat equation boundary of ∆γ(0) = 10 mm as well ∆γ ˙ (0) = −3 mm h−1 with time-varying domain. Automatica 54, 41 – 48. for the growth velocity have been introduced. As shown in Jadachowski, L., Meurer, T., Kugi, A., 2012. An ef- Figure 3, the controller ensures convergence for the error ficient implementation of backstepping observers for system e(z,t) as well as the interface deviation ∆γ(t). time-varying parabolic pdes. IFAC Proceedings Volumes 45 (2), 798 – 803, 7th Vienna International Conference 7. CONCLUSION AND OUTLOOK on Mathematical Modelling. Jurisch, M., Börner, F., Bünger, T., Eichler, S., Flade, T., Kretzer, U., Köhler, A., Stenzenberger, J., Weinert, B., In this contribution, a backstepping-based tracking control 2005. LEC- and VGF-growth of SI GaAs single crys- has been presented for a two-phase Stefan problem, occur- tals—recent developments and current issues. Journal ring in the VGF process. By utilising both system inputs, of Crystal Growth 275 (1), 283 – 291, proceedings of this approach enables the tracking of arbitrary 3 reference the 14th International Conference on Crystal Growth trajectories. Although only a linear approximation of the and the 12th International Conference on Vapor Growth error system is used, the results are promising. However, and Epitaxy. the stability proof of the modified target system for the Kang, S., Zabaras, N., 1995. Control of the freezing inter- two-phase case remains open. Moreover, for output feed- face motion in two-dimensional solidification processes back in practical applications, the proposed feedback has using the adjoint method. International Journal for Nu- to be complemented by an appropriate observer. While two merical Methods in Engineering 38 (1), 63–80. early-lumping observers are presented in (Ecklebe et al., Koga, S., Diagne, M., Krstic, M., Feb 2019. Control and 2019) by the authors for the TPSP and a late-lumping State Estimation of the One-Phase Stefan Problem via 3 The only hard restriction on the chosen references is given by the Backstepping Design. IEEE Transactions on Automatic Gevrey order condition which is rather easy to satisfy. Control 64 (2), 510–525. Koga, S., Krstic, M., May 2019. Single-Boundary Con- e˜◦(0,t)= T˜◦(0,t) − T˜◦,r(0,t)= Tm − Tm =0 (A.2) trol of the Two-Phase Stefan System. arXiv e-prints, ˜ ˜ ˜ ˜ ˜ arXiv:1905.12735. ∂z˜e˜◦ Γ◦,t = ∂z˜T◦ Γ◦,t − ∂z˜T◦,r Γ◦,t Maidi, A., Corriou, J.-P., 2016. Boundary Geometric   δ◦     Control of a Nonlinear Diffusion System with Time- = u◦(t) − ∂z˜T˜◦,r(Γ˜,t)=˜u◦(t). (A.3) Dependent Spatial Domain. Asian Journal of Control λ◦ 18 (4), 1259–1268. Appendix B. KERNEL EQUATIONS Meurer, T., Kugi, A., 2009. Tracking control for bound- ary controlled parabolic PDEs with varying parameters: Combining backstepping and differential flatness. Auto- To derive the mandatory conditions on the transformation matica 45 (5), 1182 – 1194. kernel from (23), the derivatives of (23) are taken wrt. t Petrus, B., Bentsman, J., Thomas, B., 2012. Enthalpy- and z˜. Next, they are substituted in (22a) to express based feedback control algorithms for the Stefan prob- the target system dynamics in terms of the original error lem. In: Proceedings of the IEEE Conference on Decision e˜◦ (˜z,t). Furthermore, (15) is substituted and integration and Control. pp. 7037–7042. by parts is performed to shift the spatial operators onto ˜ Petrus, B., Bentsman, J., Thomas, B. G., 2010. Feedback k(˜z,ζ,t). Finally, since the resulting equation has to hold control of the two-phase Stefan problem, with an appli- for all e˜◦ (˜z,t), one arrives at the kernel equations 2 2 cation to the continuous casting of steel. In: 49th IEEE ∂tk◦(˜z,ζ,t)= α◦ ∂z k◦(˜z,ζ,t) − ∂ζ k◦(˜z,ζ,t) Conference on Decision and Control (CDC). pp. 1731– + a◦(ζ,t)k◦(˜z,ζ,t)
(B.1a) 1736. d Petrus, B., Bentsman, J., Thomas, B. G., 2014. Applica- a◦(˜z,t) =2α◦ dz˜k◦(˜z, z,t˜ ) (B.1b) tion of Enthalpy-Based Feedback Control Methodology z˜ ¯ ¯ to the Two-Sided Stefan Problem. In: 2014 American α◦k◦(˜z, 0,t)= b◦(ζ,t)k◦(˜z,ζ,t) dζ − b◦(˜z,t) (B.1c) 0 Control Conference. pp. 1015–1020. Z z˜ Rudolph, J., Winkler, J., Woittennek, F., 2003. Flatness 0= c¯◦(ζ,t)k◦(˜z,ζ,t) dζ − c¯◦(˜z,t). (B.1d) based control of distributed parameter systems: Exam- Z0 ples and computer exercises from various technological Thus, integration of (B.1b) yields domains. Berichte aus der Steuerungs- und Regelung- 1 z˜ stechnik. Shaker Verlag, Aachen. k◦(˜z, z,t˜ )= a◦(ζ,t) dζ + k◦(0, 0,t) Rudolph, J., Winkler, J., Woittennek, F., 2004. Flatness 2α◦ Z0 based trajectory planning for two heat conduction prob- where k◦(0, 0,t)= −b◦(˜z,t)/α◦ from (B.1c) at z˜ =0. lems in crystal growth technology. e-STA (Sciences et Technologies de l’Automatique) 1 (1). Appendix C. DERIVATIVE ORDERS Stefan, J., 1891. Über die Theorie der Eisbildung, insbeson- dere über die Eisbildung im Polarmeere. Annalen der To analyse the required derivative orders for the com- Physikalischen Chemie, 269–286. putation of an arbitrary kernel element k¯i,j (t) the l-th derivative of (31) i− j−1 ∆2 1 ∂lk¯ (t)= ∂(l+1)k¯ (t) Appendix A. ERROR SYSTEM DYNAMICS t i,j 4α t n,m n=j m=0 X X l Taking the temporal derivative of (14a) and substituting l l l−d − ∂( )a n−m ∆,t ∂( )k¯ (t) (7a) for the original and reference temperature yields d t 2 t n,m d=0   ! Xj−1
∂te˜◦ (˜z,t)= ∂tT˜◦(˜z,t) − ∂tT˜◦,r(˜z,t) l ∆ l (l)¯ j+m (l−d)¯ 2 ˜ ˜ + ∂t b 2 ∆,t ∂t kj,m(t) = α◦∂z˜T◦(˜z,t) +γ ˙ (t)∂z˜T◦(˜z,t) α d m=0 d=0 2   − α◦∂ T˜◦, (˜z,t) − γ˙ (t)∂zT˜◦, (˜z,t) X X i−
z˜ r r ˜ r 1 ∆ 1 − ∂l¯b (j∆,t)+ ∂la n ∆,t (C.1) which by substituting (14a), derived wrt. z˜ twice, becomes: α t 4α t 2 n=j 2 ˜ ˜ X
= α◦∂z˜e˜◦ (˜z,t) +γ ˙ (t)∂z˜T◦(˜z,t) − γ˙r(t)∂z˜T◦,r(˜z,t). is examined. As it can be seen, for every derivative in ˜ (31) that is to be eliminated, derivatives of the kernel Furthermore, replacing γ˙ (t) and T◦(˜z,t) via (14a) and elements below and left of one order higher, as well as (14b), respectively, gives derivatives of the provided functions a(˜z,t) and ¯b(˜z,t) 2 ˜ with the same order are introduced (cf. shaded areas in ∂te˜◦ (˜z,t)= α◦∂z˜e˜◦ (˜z,t) − γ˙ r(t)∂z˜T◦,r(˜z,t) Figure 2). Therefore, the computation of k¯i,j (t) demands a ˜ + ∆γ ˙ (t) +γ ˙ r(t) ∂z˜e˜◦ (˜z,t)+ ∂z˜T◦,r(˜z,t) derivative order of min{i, j} = j for these functions. Hence, for the computation of the complete kernel, dmax = Nσ −1 and thus    ¯ derivatives are required as the kernel element kNσ,Nσ (t) is 2 already given by (33). ∂te˜◦ (˜z,t)= α◦∂z˜e˜◦ (˜z,t)+ ∆γ ˙ (t)∂z˜ +γ ˙ r(t)∂z˜ e˜◦ (˜z,t) ˜ + ∆γ ˙ (t)∂z˜T◦,r(˜z,t ).
(A.1) Finally, examining the boundaries yields