betaFeedback

WorksPhysics and Astronomy

Data-driven learning of the generalized Langevin equation with state-dependent memory

Abstract

We present a data-driven method to learn stochastic reduced models of complex systems that retain a state-dependent memory beyond the standard generalized Langevin equation (GLE) with a homogeneous kernel. The constructed model naturally encodes the heterogeneous energy dissipation by jointly learning a set of state features and the non-Markovian coupling among the features. Numerical results demonstrate the limitation of the standard GLE and the essential role of the broadly overlooked state-dependency nature in predicting molecule kinetics related to conformation relaxation and transition.

Introduction

Predicting the collective behavior of complex multiscale systems is often centered around projecting the full-dimensional dynamics onto a set of resolved variables. However, an accurate construction of such a reduced model remains a practical challenge for real applications such as molecular modeling. While model reduction frameworks such as the Koopman operator [1] and the Mori-Zwanzig projection formalism [2, 3] enable us to write down the dynamic equations in terms of the resolved variables, the reduced model generally becomes non-Markovian with a memory term that may further depend on the resolved variables; the direct numerical evaluation involves solving the expensive full-dimensional orthogonal dynamics. In practice, one common approximation is to ignore such state-dependency; the reduced model is simplified as the standard generalized Langevin equation (GLE) [4] with a memory kernel that only depends on time. Several approaches [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20] have been developed to construct the memory kernel such that certain dynamic properties (e.g., the two-point correlations) can be properly reproduced. Despite its broad application, the validity of the standard GLE for real multiscale systems remains less understood [21, 22].

Intuitively, the above model reduction problem is somewhat analogous to hiking on a mountain where the landscape map and the path roughness represent the free energy and the memory term, respectively. In general, we should not expect homogeneous path roughness at the different locations (e.g., the valleys and the ridges), which, conversely, needs to be inferred from the hiking records. Indeed, studies based on full molecular dynamics (MD) simulations [23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33] and sophisticated projection operator construction [34, 35, 36, 37, 38, 39, 40, 41, 42] show that the extracted memory term can exhibit a pronounced state-dependent nature where the implications for the collective behaviors remain under-explored. For extensive MD systems, a recent study [43] on reduced modeling of polymer melt shows that the heterogeneous inter-molecular energy dissipation (i.e., the memory) can be crucial for transport on the hydrodynamic scale. However, for canonical non-extensive problems such as biomolecule kinetics, a quantitative understanding of the effect of state-dependent memory arising from intra-molecular interactions remains an open problem. Several recent works [10, 13, 18, 44, 45, 46, 47, 48] investigate the non-Markovian implication for transition dynamics based on the standard GLE. While elegant semi-analytical studies [49, 50, 51, 52, 53, 54, 55, 56, 57, 58] on idealized 1D double-well potential provide theoretical insights into the state-dependent nature, accurately quantifying such effect on molecule transition dynamics relies on accurate construction and efficient simulation of a reduced model beyond the standard GLE.

This work presents a data-driven approach for learning a stochastic reduced model that retains a state-dependent memory for non-extensive systems. Instead of dealing with the orthogonal dynamics [6, 40, 43], the training only relies on the trajectory samples and the full model is unnecessarily known explicitly. The main idea is to seek a generalized representation of the memory as the composition of a set of state-dependent features, which encodes the coupling between the resolved and unresolved variables and can be learned using three-point correlation functions. Efficient training is achieved by constructing the encoder functions using a set of sparse bases, whose correlations can be efficiently pre-computed. Coherent noise can be introduced that ensures a consistent invariant distribution. The present method enables us to probe the open problem of the effect of state-dependent memory on molecular kinetics. Numerical results show that the broadly overlooked state-dependency can play a crucial role. In particular, the standard GLE could be insufficient to capture the collective properties such as the transition rate distribution, which, fortunately, can be reproduced by the present model.

Reduced model construction

Let (q,p)∈R2m(\mathbf{q}, \mathbf{p}) \in\mathbb{R}^{2m} represent the resolved variables of a high-dimensional Hamiltonian system, where q\mathbf{q} denotes the coarse-grained (CG) coordinates as a function of the position variables of the full model, and p\mathbf{p} denotes the CG momenta. Following the Zwanzig’s projection formalism [4], the reduced dynamics takes the form

q˙=M−1p,p˙=−∇U(q)−∫0tK(q(τ),t−τ)v(τ) dτ+Rt,(1)\dot{\mathbf{q}} = \mathbf{M}^{-1}\mathbf{p}, \qquad \dot{\mathbf{p}} = -\nabla U(\mathbf{q}) - \int_{0}^{t} \mathbf{K}(\mathbf{q}(\tau),t-\tau)\mathbf{v}(\tau)\,\mathrm{d}\tau+ \mathbf{R}_{t}, \tag*{(1)}

where M\mathbf{M} is the mass matrix, U(q)U(\mathbf{q}) is the free energy, v:=q˙\mathbf{v} := \dot{\mathbf{q}} is the velocity, K(q,t)\mathbf{K}(\mathbf{q},t) is the memory, and Rt\mathbf{R}_{t} is the noise whose covariance function is related to the memory following the second fluctuation-dissipation theorem [40]. Before proceeding to the construction of K(q,t)\mathbf{K}(\mathbf{q},t), we note that M\mathbf{M} generally depends on q\mathbf{q}. In this study, we focus on the effect of the state-dependent memory; the current choice of q\mathbf{q} leads to a constant mass matrix (see Refs. [13, 41] and Appendix for further discussion). Also, the construction of the free energy U(q)U(\mathbf{q}) can be nontrivial; several canonical methods based on enhanced sampling [59, 60, 61, 62] and temperature acceleration [63, 64, 65, 66] have been developed. Here, we assume U(q)U(\mathbf{q}) is known a priori.

Instead of rigorously constructing K(q,t)\mathbf{K}(\mathbf{q},t) from the full model, we ask the question of which forms of K\mathbf{K} can generate a memory effect. One common approach is to embed the memory in a larger Markovian dynamics with a set of auxiliary variables. An essential observation is that the memory term can be generally written as

K(q(τ),t−τ)≈C+∘exp⁡((t−τ)Laux)∘C−,(2)\mathbf{K}(\mathbf{q}(\tau),t-\tau) \approx\mathcal{C}^{+} \circ\exp\left((t-\tau)\mathcal{L}_{\mathrm{aux}}\right) \circ\mathcal{C}^{-}, \tag*{(2)}

where Laux\mathcal{L}_{\mathrm{aux}} is the Liouville operator corresponding to the auxiliary dynamics and C±\mathcal{C}^{\pm} are channels representing the coupling of the resolved and auxiliary variables. As a special case, if the coupling and the auxiliary dynamics take a linear form, the embedded memory recovers the standard GLE kernel, i.e., K(q,t)=K(t)\mathbf{K}(\mathbf{q},t)=\mathbf{K}(t) (e.g., see Refs. [19, 67]). Therefore, to construct the reduced model beyond the standard GLE, the coupling channels need to properly retain certain kinds of state-dependency nature. This motivates us to represent C±\mathcal{C}^{\pm} by seeking a set of state-dependent features ϕ(q)=[ϕ1(q),…,ϕn(q)]\boldsymbol{\phi}(\mathbf{q})=[\boldsymbol{\phi}_1(\mathbf{q}),\ldots,\boldsymbol{\phi}_n(\mathbf{q})], where ϕ:Rm→Rn×m\boldsymbol{\phi}:\mathbb{R}^{m}\to\mathbb{R}^{n\times m} essentially encode the nonlinear coupling between the resolved and unresolved variables and the detailed form will be specified later. exp⁡(tLaux)\exp(t\mathcal{L}_{\mathrm{aux}}) induces the non-Markovian interactions among the features with a time lag of tt characterized by a kernel function, i.e.,

C+exp⁡((t−τ)Laux)∘C−=ϕ(q(t))TΘ(t−τ)ϕ(q(τ))\mathcal{C}^{+}\exp((t-\tau)\mathcal{L}_{\mathrm{aux}})\circ\mathcal{C}^{-}=\boldsymbol{\phi}(\mathbf{q}(t))^{T}\boldsymbol{\Theta}(t-\tau)\boldsymbol{\phi}(\mathbf{q}(\tau)), where Θ:R+→Rn×n\boldsymbol{\Theta}:\mathbb{R}^{+}\to\mathbb{R}^{n\times n} and component Θij(t−τ)\Theta_{ij}(t-\tau) represents the dissipation between features ϕi(q(t))\boldsymbol{\phi}_i(\mathbf{q}(t)) and ϕj(q(τ))\boldsymbol{\phi}_j(\mathbf{q}(\tau)). In the remainder of this work, we use ϕt\boldsymbol{\phi}_t to denote ϕ(q(t))\boldsymbol{\phi}(\mathbf{q}(t)).

Accordingly, reduced dynamics (1) can be modeled by

q˙=M−1p,\dot{\mathbf{q}}=\mathbf{M}^{-1}\mathbf{p},
p˙=−∇U(q)−∫0tϕtTΘ(t−τ)ϕτv(τ) dτ+Rt,(3)\dot{\mathbf{p}}=-\nabla U(\mathbf{q})-\int_{0}^{t}\boldsymbol{\phi}_t^{T}\boldsymbol{\Theta}(t-\tau)\boldsymbol{\phi}_{\tau}\mathbf{v}(\tau)\,\mathrm{d}\tau+\mathcal{R}_t, \tag*{(3)}

where encoders {ϕi(q)}i=1n\{\boldsymbol{\phi}_i(\mathbf{q})\}_{i=1}^{n} and kernel Θ(t)\boldsymbol{\Theta}(t) need to be determined. As a special case, at the Markovian limit Θ(t)∝δ(t)\boldsymbol{\Theta}(t)\propto\delta(t), Eq. (3) recovers the Langevin dynamics and the quadratic form ϕTϕ\boldsymbol{\phi}^{T}\boldsymbol{\phi} ensures positive energy dissipation. Also, by choosing Θ(t)\boldsymbol{\Theta}(t) to be diagonal with individual components corresponding to certain frequency modes, Eq. (3) reduces to the heat bath model [35] with a nonlinear coupling of bath coordinates. On the other hand, the present model enables an adaptive choice of the number of spatial features and a more general form of Θ(t)\Theta(t) with the off-diagonal components capturing the non-Markovian coupling among the features, which turns out to be crucial for reproducing the collective dynamics.

We emphasize that Eq. (3) should not be viewed as a direct approximation of Zwanzig’s projection formalism. Rather, it serves as a reduced model that faithfully retains the state-dependent memory effect. To construct the model, we represent encoders {ϕi(q)}i=1n\{\phi_i(\mathbf{q})\}_{i=1}^{n} and kernel Θ(t)\Theta(t) in form of

ϕi(q)=HiTψ(q),Θ(t)=e−αt∑k=0NωΘ^kcos⁡(ωkt),(4)\begin{aligned} \phi_i(\mathbf{q}) &= \mathbf{H}_i^{T}\psi(\mathbf{q}),\\ \Theta(t) &= e^{-\alpha t}\sum_{k=0}^{N_{\omega}}\widehat{\Theta}_k\cos(\omega_k t), \tag*{(4)} \end{aligned}

where ψ(q)=[ψ1(q),⋯ ,ψNb(q)]\psi(\mathbf{q})=[\psi_1(\mathbf{q}),\cdots,\psi_{N_b}(\mathbf{q})] is a set of sparse bases and H=[H1T,⋯ ,HnT]\mathbf{H}=[\mathbf{H}_1^T,\cdots,\mathbf{H}_n^T] are trainable coefficients to be determined. In this work, we choose the piece-wise linear bases such that the correlation between ϕi\phi_i and ϕj\phi_j can be efficiently evaluated; other choices such as localized kernel can be also used. Θ(t)\Theta(t) needs to preserve positive semi-definiteness. Hence, we represent Fourier modes Θ^k=ΓkΓkT\widehat{\Theta}_k=\Gamma_k\Gamma_k^T, where Γk∈Rn×n\Gamma_k\in\mathbb{R}^{n\times n} is a low-triangular matrix to be determined along with α≥0\alpha\geq0. For the present study, the full dynamics is reversible; α\alpha approaches 0 and essentially serves as a regularization parameter. We note that Θ(t)\Theta(t) in the form of Eq. (17) can be further generalized by introducing an anti-symmetry part and refer to Appendix for further discussion.

To learn the reduced model (3), we need to choose appropriate metrics such that the state-dependent non-Markovian nature can be manifested. While auto-correlation functions such as cvv(t)=⟨v(t)v(0)T⟩c_{vv}(t)=\langle\mathbf{v}(t)\mathbf{v}(0)^T\rangle merely characterize the overall memory effect, a crucial observation is that the correlations conditional with different initial state q0\mathbf{q}_0 further depends on the local energy dissipation and therefore naturally encodes the signatures of the heterogeneous memory effect. Accordingly, we right-multiply the second equation of Eq. (3) by v(0)\mathbf{v}(0) and take the conditional expectation on q0=q∗\mathbf{q}_0=\mathbf{q}^*, i.e.,

g(t;q∗)=∫0t⟨ϕtTΘt−τϕτvτv0T∣q0=q∗⟩ dτ=∫0tTr⁡⟨Θt−τHψτvτv0TψtTHT∣q0=q∗⟩ dτ=∫0tTr⁡[Θt−τHCψ,ψ(t,τ;q∗)HT] dτ,\begin{aligned} \mathbf{g}(t;\mathbf{q}^*) &= \int_0^t\left\langle\phi_t^T\Theta_{t-\tau}\phi_\tau\mathbf{v}_\tau\mathbf{v}_0^T\mid\mathbf{q}_0=\mathbf{q}^*\right\rangle\,\mathrm{d}\tau\\ &=\int_0^t\operatorname{Tr}\left\langle\Theta_{t-\tau}\mathbf{H}\psi_\tau\mathbf{v}_\tau\mathbf{v}_0^T\psi_t^T\mathbf{H}^T\mid\mathbf{q}_0=\mathbf{q}^*\right\rangle\,\mathrm{d}\tau\\ &=\int_0^t\operatorname{Tr}\left[\Theta_{t-\tau}\mathbf{H}C_{\psi,\psi}(t,\tau;\mathbf{q}^*)\mathbf{H}^T\right]\,\mathrm{d}\tau, \end{aligned}

where g(t;q∗):=⟨[p˙t+∇U(qt)]v0T∣q0=q∗⟩\mathbf{g}(t;\mathbf{q}^*):=\left\langle[\dot{\mathbf{p}}_t+\nabla U(\mathbf{q}_t)]\mathbf{v}_0^T\mid\mathbf{q}_0=\mathbf{q}^*\right\rangle and Cψ,ψ(t,τ;q∗):=⟨ψτvτv0TψtT∣q0=q∗⟩C_{\psi,\psi}(t,\tau;\mathbf{q}^*):=\left\langle\psi_\tau\mathbf{v}_\tau\mathbf{v}_0^T\psi_t^T\mid\mathbf{q}_0=\mathbf{q}^*\right\rangle is a three-point correlation characterizing the coupling among the bases. Since ψ(q)\psi(\mathbf{q}) is sparse, ψτψτT\psi_{\tau}\psi_{\tau}^{T} can be evaluated with O(1)O(1) complexity and hence Cψ,ψ(t,τ;q∗)\mathbf{C}_{\psi,\psi}(t,\tau;\mathbf{q}^{*}) can be efficiently pre-computed. Accordingly, we can establish the training of the reduced model in terms of coefficients H\mathbf{H} for encoders ϕ(q)\phi(\mathbf{q}) as well as matrices {Γk}k=1Nω\{\mathbf{\Gamma}_{k}\}_{k=1}^{N_{\omega}} and α\alpha for kernel Θ(t)\mathbf{\Theta}(t) by minimizing the empirical loss

L=∑l=1Nq∑k=1Nt∥g~(tk;q(l))−g(tk;q(l))∥2,L=\sum_{l=1}^{N_{q}}\sum_{k=1}^{N_{t}}\left\|\widetilde{\mathbf{g}}(t_{k};\mathbf{q}^{(l)})-\mathbf{g}(t_{k};\mathbf{q}^{(l)})\right\|^{2},
g~(tk;q(l))=∑j=1kTr⁡[Θ(tk−tj)HCψ,ψ(tk,tj;q(l))HT]δt,\widetilde{\mathbf{g}}(t_{k};\mathbf{q}^{(l)})=\sum_{j=1}^{k}\operatorname{Tr}\left[\mathbf{\Theta}(t_{k}-t_{j})\mathbf{H}\mathbf{C}_{\psi,\psi}(t_{k},t_{j};\mathbf{q}^{(l)})\mathbf{H}^{T}\right]\delta t,

where ll and kk correspond to various initial states and discrete time, respectively. δt\delta t is the time step of training samples. g~(⋅)\widetilde{\mathbf{g}}(\cdot) represents the prediction by the reduced model which depends on the trainable variables and the pre-computed correlation Cψ,ψ\mathbf{C}_{\psi,\psi}.

To simulate the reduced model (3) (C1) on t∈[0,T]t\in[0,T], we generate coherent noise Rt=ϕtTR~(t)\mathbf{R}_{t}=\phi_{t}^{T}\widetilde{\mathbf{R}}(t), where R~:R+→Rn\widetilde{\mathbf{R}}:\mathbb{R}^{+}\to\mathbb{R}^{n} is a Gaussian random process. Specifically, we can show that by choosing ⟨R~(t)R~(τ)T⟩=kBTe−α(t−τ)Θ(t−τ)\left\langle\widetilde{\mathbf{R}}(t)\widetilde{\mathbf{R}}(\tau)^{T}\right\rangle=k_{B}Te^{-\alpha(t-\tau)}\mathbf{\Theta}(t-\tau), the reduced model retains a consistent equilibrium density, i.e., ρeq(q,p)∝exp⁡{−β[U(q)+12pTM−1p]}\rho_{\mathrm{eq}}(\mathbf{q},\mathbf{p})\propto\exp\left\{-\beta\left[U(\mathbf{q})+\frac{1}{2}\mathbf{p}^{T}\mathbf{M}^{-1}\mathbf{p}\right]\right\} (see proof in Appendix). Accordingly, we can generate discrete samples {R~(ti)}i=0N\{\widetilde{\mathbf{R}}(t_{i})\}_{i=0}^{N} by

R~(ti)=β−1/2∑k=02NΘ~k1/2[cos⁡(ωkti)ξk+sin⁡(ωkti)ηk],(5)\widetilde{\mathbf{R}}(t_{i})=\beta^{-1/2}\sum_{k=0}^{2N}\widetilde{\mathbf{\Theta}}_{k}^{1/2}\left[\cos(\omega_{k}t_{i})\boldsymbol{\xi}_{k}+\sin(\omega_{k}t_{i})\boldsymbol{\eta}_{k}\right], \tag*{(5)}

where Θ~k\widetilde{\mathbf{\Theta}}_{k} are the Fourier (essentially cosine) modes of e−α∣t∣Θ(t)e^{-\alpha|t|}\mathbf{\Theta}(t) on [−T,T][-T,T] (see Refs. [68, 69] and Appendix for the analytical form); ξk\boldsymbol{\xi}_{k} and ηk\boldsymbol{\eta}_{k} are independent Gaussian random vectors. In practice, R~(t)\widetilde{\mathbf{R}}(t) by Eq. (5) can be generated using FFT [70] and the convolution term ∫0tϕτTΘ(t−τ)ϕτv(τ) dτ\int_{0}^{t}\phi_{\tau}^{T}\mathbf{\Theta}(t-\tau)\phi_{\tau}\mathbf{v}(\tau)\,\mathrm{d}\tau in Eq. (3) can be efficiently evaluated using the fast convolution algorithm [71], both of which only require O(Nlog⁡N)O(N\log N) complexity.

Numerical results

The present reduced model enables us to systematically investigate the open problem of the effect of state-dependent memory on the collective dynamics of complex systems such as molecule kinetics. In this work, we consider the molecule benzyl bromide in an aqueous environment. The full MD system consists of one benzyl bromide molecule and 2400 water molecules with the periodic boundary condition. The isothermal-isobaric thermostat [72] is used to equilibrate the system at 298K and 1 bar and a canonical ensemble with a Nosé-Hoover thermostat [73, 74] is used for the production stage. The resolved variable q\mathbf{q} characterizes the interplay between the substituent and the benzene group and is defined as the distance between the bromine atom and the ipso-carbon atom.

Let us start with the standard GLE by setting features ϕ(q)≡I\phi(\mathbf{q}) \equiv\mathbf{I} in Eq. (3). We right-multiply q(0)\mathbf{q}(0) (or v(0)\mathbf{v}(0)) to Eq. (3) and compute the correlation functions, i.e., h(t)=∫0tΘ(t−τ)cvq(τ) dτh(t) = \int_{0}^{t} \Theta(t-\tau)c_{v_q}(\tau)\,\mathrm{d}\tau, where h(t)=⟨[p˙t+∇U(qt)]q0T⟩h(t) = \left\langle[\dot{\mathbf{p}}_{t}+\nabla U(\mathbf{q}_{t})]\mathbf{q}_{0}^{\mathsf{T}}\right\rangle. The standard GLE kernel Θ(t)\boldsymbol{\Theta}(t) (i.e., K(t)\mathbf{K}(t) in Eq. (1)) can be computed using the Fourier transform of the integral equation. If the reduced dynamics (1) can be simplified as the standard GLE, then cvq(t)c_{v_q}(t) should be accurately reproduced. Figure 1 shows the prediction of cvq(t)c_{v_q}(t) from the standard GLE and the full MD model. The apparent deviations imply non-negligible state-dependency. To further probe this effect, we compute h(t;q∗)h(t;\mathbf{q}^{*}) conditional with different initial states q∗\mathbf{q}^{*}. Unlike a unified correlation predicted by the standard GLE, the large dispersion reveals the heterogeneous nature of the energy dissipation process.

Figure 1. Correlation functions predicted by the standard GLE and the full MD: (a) Overall cvq(t)c_{v_q}(t) and (b) conditional h(t;q∗)h(t;\mathbf{q}^{*}) with q∗\mathbf{q}^{*} representing various initial states (gray lines), including two local minima and the saddle point (see inset of Figure 2(a)). The large dispersion implies the limitation of the standard GLE, which predicts a single curve for h(t)h(t) in short time.

To capture the state-dependent memory, we train the present model (3) with a different number of features. Figure 2(a-Figure 2b) shows the obtained encoder ϕ(⋅)\phi(\cdot) using one feature and Θ(t)\boldsymbol{\Theta}(t) is scaled such that Θ(0)=1\boldsymbol{\Theta}(0)=1. We can see that ϕ\phi exhibits apparent deviation from a uniform distribution. In particular, it shows a peak value near the saddle point q=3.65q = 3.65, implying a larger effective friction near the regime. This result supports a similar assumption in earlier semi-analytical studies (e.g., see Ref. [55]) on improving Kramers’ rate theory [75]. Also, it explains the short-time dispersion shown in Figure 1, where h(t;q∗)h(t; q^{*}) with q∗q^{*} near the saddle point decays faster than the ones near the local minima. Furthermore, we can train the reduced model using multiple features. Figure 2(c-Figure 2d) shows the obtained encoders {ϕi(⋅)}i=1n\{\phi_i(\cdot)\}_{i=1}^{n} with n=4n = 4 and the diagonal components of Θ(t)\boldsymbol{\Theta}(t). Compared with the case of n=1n = 1, the larger variation of ϕi\phi_i enables a better representation of the state-dependent memory.

Figure 2. The state features ϕ\phi and diagonal components of the matrix-valued kernel Θ(t)\boldsymbol{\Theta}(t) for the present model with state-dependent memory (SD-GLE) trained using (a-b) one feature and (c-d) four features. Inset plots: (a) probability density function (PDF) of qq, where ϕ(q)\phi(q) near the saddle point shows a pronounced peak; (b) Fourier modes of Θ(t)\boldsymbol{\Theta}(t) for the present model and standard GLE.

With the constructed model, we examine the conditional correlations cvq(t;q∗)c_{vq}(t; q^{*}) and cvv(t;q∗)c_{vv}(t; q^{*}). As shown in Figure 3, for both the local minima and the saddle point, the predictions of the present model using four features show good agreement with the full MD results. In contrast, the predictions of the standard GLE show apparent deviations for q∗q^{*} as the saddle point. Also, we note that the present model using one feature shows improved short-time predictions but remains insufficient for long-time correlations. This reveals the complex global variation of the memory term, which can not be simply represented by a single feature as a state-dependent re-scaling of the kernel function; the non-Markovian coupling among multiple features is crucial to capture the heterogeneous energy dissipation over the full space.

Figure 3. The overall and conditional correlation functions predicted by the full MD and various reduced models for two local minimal and the saddle point: (a-b) cvqc_{vq} and (c-d) cvvc_{vv}.

Finally, we examine the collective behavior related to molecule kinetics. Figure 4(a) shows the position correlation cqq(t)c_{qq}(t) characterizing the relaxation of the molecule conformation. Compared with the MD results, the standard GLE shows a significant underestimation of the overall relaxation time. This discrepancy is possibly due to the larger effective friction near the saddle point (see Figure 2(a)), which essentially dampens the transition between the two local minima. The standard GLE overlooked such state dependency and therefore yields a faster relaxation. This limitation is consistently reflected in the distribution of the transition time. As shown in Figure 4(b), the standard GLE predicts a larger probability for the short transition time, indicating a smaller overall friction than the local (i.e., saddle point) value. Fortunately, the heterogeneous non-Markovianity can be faithfully retained in the present model. In particular, the constructed model using a single feature yields a better prediction than the standard GLE. As we increase to four features, the predictions recover the MD results.

Figure 4. Collective molecule behaviors predicted by the full MD and the various reduced models: (a) overall conformation relaxation and (b) distribution of the transition time between the two local minima.

Summary

To conclude, we present a data-driven approach for learning a stochastic reduced model beyond the standard GLE, where the complex state-dependent memory can be naturally encoded in the non-Markovian interactions among a set of features in terms of the resolved variables. The training does not rely on the explicit knowledge of the full model and only utilizes the trajectory samples, where the three-point correlations can be efficiently pre-computed. Numerical results of a molecule system demonstrate the crucial role of the state-dependent non-Markovianity on collective behavior, where the standard GLE shows limitations due to the over-simplified assumption of a homogeneous memory kernel. In contrast, the present model accurately predicts the molecule kinetics including the transition time distribution, and paves the way toward predictive modeling of the collective functional properties and rare events [76] of complex multi-scale biomolecule and material systems.

Acknowledgments

We acknowledge helpful discussions from Zhaopeng Hao and Peiyuan Gao. The work is supported in part by the National Science Foundation under Grant DMS-2110981 and the ACCESS program through allocation MTH210005.

Mass matrix of the reduced model

In this study, we focus on the effect of the state-dependent non-Markovian memory on the collective behavior of complex systems. We choose the coarse-grained resolved variables such that the corresponding mass matrix is a constant. On the other hand, the mass matrix should further depend on the resolved variables for the general cases, and we refer to Refs. [13, 41] for further discussions and the reduced dynamics with position-dependent mass. Specifically, we define q=∥Q1−Q2∥q = \lVert\mathbf{Q}_{1} - \mathbf{Q}_{2} \rVert, where Q1\mathbf{Q}_{1} and Q2\mathbf{Q}_{2} are the atom coordinates of the full model (see Figure 5 and Sec. E for details). Accordingly, we have q˙=Q12TQ˙12/q\dot{q} = \mathbf{Q}_{12}^{T}\dot{\mathbf{Q}}_{12}/q and its covariance follows

A sketch of the molecule benzyl bromide

Figure 5. A sketch of the molecule benzyl bromide. The resolved variable is defined as the distance between the bromine atom and the ipso-carbon atom.

⟨q˙q˙⟩=⟨1q2Q12TQ˙12Q˙12TQ12⟩=⟨1q2Tr⁡[(Q12Q12T)(Q˙12Q˙12T)]⟩=⟨1q2Tr⁡(Q12Q12T)⟩(M1−1+M2−1)kBT=(M1−1+M2−1)kBT,(6)\begin{aligned} \langle\dot{q}\dot{q} \rangle&= \left\langle\frac{1}{q^{2}}\mathbf{Q}_{12}^{T}\dot{\mathbf{Q}}_{12}\dot{\mathbf{Q}}_{12}^{T}\mathbf{Q}_{12} \right\rangle\\ &= \left\langle\frac{1}{q^{2}}\operatorname{Tr}\left[(\mathbf{Q}_{12}\mathbf{Q}_{12}^{T})(\dot{\mathbf{Q}}_{12}\dot{\mathbf{Q}}_{12}^{T})\right] \right\rangle\\ &= \left\langle\frac{1}{q^{2}}\operatorname{Tr}(\mathbf{Q}_{12}\mathbf{Q}_{12}^{T}) \right\rangle\left(M_{1}^{-1} + M_{2}^{-1}\right) k_{B}T \\ &= \left(M_{1}^{-1} + M_{2}^{-1}\right) k_{B}T, \tag*{(6)} \end{aligned}

where M1M_{1} and M2M_{2} represent the mass of two atoms and we have used the fact that the distribution of Q12\mathbf{Q}_{12} and Q˙12\dot{\mathbf{Q}}_{12} are independent. Therefore, the mass matrix of qq is a constant M≡M1M2/(M1+M2)M \equiv M_{1}M_{2}/(M_{1} + M_{2}).

Coherent noise and invariant density of the reduced model

We construct the reduced model

q˙=M−1p,p˙=−∇U(q)−∫0tϕtTΘ(t−τ)ϕτv(τ) dτ+Rt,(7)\dot{\mathbf{q}}=\mathbf{M}^{-1}\mathbf{p}, \qquad \dot{\mathbf{p}}=-\nabla U(\mathbf{q})-\int_{0}^{t}\boldsymbol{\phi}_{t}^{T}\boldsymbol{\Theta}(t-\tau)\boldsymbol{\phi}_{\tau}\mathbf{v}(\tau)\,\mathrm{d}\tau+\mathcal{R}_{t}, \tag*{(7)}

where ϕt=[ϕ1(qt),…,ϕn(qt)]\boldsymbol{\phi}_{t}=[\phi_{1}(\mathbf{q}_{t}),\ldots,\phi_{n}(\mathbf{q}_{t})] and Θ(t)\boldsymbol{\Theta}(t) represent the spatial features and the kernel to be learned. In particular, Θ(t)\boldsymbol{\Theta}(t) is directly constructed in the Fourier space, i.e., Θ(t)=e−αt∑k=0NωΘ^kcos⁡(ωkt)\boldsymbol{\Theta}(t)=e^{-\alpha t}\sum_{k=0}^{N_{\omega}}\widehat{\boldsymbol{\Theta}}_{k}\cos(\omega_{k}t), where ωk=2πTck\omega_{k}=\frac{2\pi}{T_{c}}k and TcT_{c} is the time domain cut-off of the kernel. We note that e−αte^{-\alpha t} should not be viewed as the bases to approximate Θ(t)\boldsymbol{\Theta}(t) (e.g., {e−αitcos⁡(βit),e−αitsin⁡(βit)}i=1Nα\{e^{-\alpha_{i}t}\cos(\beta_{i}t),e^{-\alpha_{i}t}\sin(\beta_{i}t)\}_{i=1}^{N_{\alpha}}; see Refs. [10, 13]). Rather, Θ(t)\boldsymbol{\Theta}(t) is mainly characterized by the Fourier series expansion on [0,T][0,T], and the exponential term e−αte^{-\alpha t} is essentially a regularization term to eliminate the periodicity while maintaining the semi-positive definiteness condition.

For the fluctuation term Rt\mathcal{R}_{t}, we represent it as a noise in the form of Rt=ϕtTR~(t)\mathcal{R}_{t}=\boldsymbol{\phi}_{t}^{T}\widetilde{\mathbf{R}}(t), where R~(t)\widetilde{\mathbf{R}}(t) is a Gaussian random process whose covariance function determined by Θ(t)\boldsymbol{\Theta}(t), i.e., ⟨R~(t)R~(τ)T⟩=kBTe−α(t−τ)Θ(t−τ)\langle\widetilde{\mathbf{R}}(t)\widetilde{\mathbf{R}}(\tau)^{T}\rangle=k_{B}Te^{-\alpha(t-\tau)}\boldsymbol{\Theta}(t-\tau). This choice avoids dealing with the orthogonal dynamics to calculate the fluctuation term. Furthermore, we can show that this choice enables the reduced model to retain a consistent invariant density function.

Proposition B.1. For reduced model (B1) with Θ(t)=e−αt∑k=0NωΘ^kcos⁡(ωkt)\boldsymbol{\Theta}(t)=e^{-\alpha t}\sum_{k=0}^{N_{\omega}}\widehat{\boldsymbol{\Theta}}_{k}\cos(\omega_{k}t), by choosing the fluctuation term Rt=ϕtTR~(t)\mathcal{R}_{t}=\boldsymbol{\phi}_{t}^{T}\widetilde{\mathbf{R}}(t), where R~(t)\widetilde{\mathbf{R}}(t) is a Gaussian random process satisfying

⟨R~(t)R~(τ)T⟩=kBTe−α(t−τ)Θ(t−τ),(8)\langle\widetilde{\mathbf{R}}(t)\widetilde{\mathbf{R}}(\tau)^{T}\rangle=k_{B}Te^{-\alpha(t-\tau)}\boldsymbol{\Theta}(t-\tau), \tag*{(8)}

the reduced model has an invariant distribution

ρeq(q,p)∝exp⁡{−[U(q)+pTM−1p/2]/kBT}.(9)\rho_{\mathrm{eq}}(\mathbf{q},\mathbf{p})\propto\exp\left\{-\left[U(\mathbf{q})+\mathbf{p}^{T}\mathbf{M}^{-1}\mathbf{p}/2\right]/k_{B}T\right\}. \tag*{(9)}

Proof. Let us introduce auxiliary variables

zk,1=−∫0te−α(t−τ)Γkcos⁡(ωk(t−τ))ϕτv(τ) dτ+Rk,1(t),zk,2=−∫0te−α(t−τ)Γksin⁡(ωk(t−τ))ϕτv(τ) dτ+Rk,2(t),(10)\begin{aligned} \mathbf{z}_{k,1}&=-\int_{0}^{t}e^{-\alpha(t-\tau)}\boldsymbol{\Gamma}_{k}\cos(\omega_{k}(t-\tau))\boldsymbol{\phi}_{\tau}\mathbf{v}(\tau)\,\mathrm{d}\tau+\mathbf{R}_{k,1}(t),\\ \mathbf{z}_{k,2}&=-\int_{0}^{t}e^{-\alpha(t-\tau)}\boldsymbol{\Gamma}_{k}\sin(\omega_{k}(t-\tau))\boldsymbol{\phi}_{\tau}\mathbf{v}(\tau)\,\mathrm{d}\tau+\mathbf{R}_{k,2}(t), \tag*{(10)} \end{aligned}

where ΓkTΓk=Θ^k\boldsymbol{\Gamma}_{k}^{T}\boldsymbol{\Gamma}_{k}=\widehat{\boldsymbol{\Theta}}_{k} and Rk,1(t)\mathbf{R}_{k,1}(t) is a Gaussian random process satisfying

⟨Rj,1(t)Rk,1(τ)T⟩=kBTδjke−α(t−τ)cos⁡(ωk(t−τ)),(11)\left\langle\mathbf{R}_{j,1}(t)\mathbf{R}_{k,1}(\tau)^{T}\right\rangle=k_{B}T\delta_{jk}e^{-\alpha(t-\tau)}\cos(\omega_{k}(t-\tau)), \tag*{(11)}

where δjk\delta_{jk} is the Kronecker delta. Accordingly, the second equation of Eq. (7) can be written as

p˙=−∇U(q)+ϕ(q)T∑kΓkTzk,1,(12)\dot{\mathbf{p}}=-\nabla U(\mathbf{q})+\phi(\mathbf{q})^{T}\sum_k\mathbf{\Gamma}_k^{T}\mathbf{z}_{k,1}, \tag*{(12)}

and Rk,2(t)\mathbf{R}_{k,2}(t) will be specified later.

Let zk=[zk,1,zk,2]\mathbf{z}_k=[\mathbf{z}_{k,1},\mathbf{z}_{k,2}] and Rk=[Rk,1,Rk,2]\mathbf{R}_k=[\mathbf{R}_{k,1},\mathbf{R}_{k,2}], we can rewrite Eq. (10) by

zk=−∫0te−α(t−τ)(cos⁡(ωk(t−τ))Isin⁡(ωk(t−τ))I−sin⁡(ωk(t−τ))Icos⁡(ωk(t−τ))I)(ΓkϕTv(τ)0) dτ+Rk(t)=−∫0texp⁡[(−αIωkI−ωkI−αI)(t−τ)](ΓkϕTv(τ)0) dτ+Rk(t).(13)\begin{aligned} \mathbf{z}_k={}&-\int_0^t e^{-\alpha(t-\tau)} \begin{pmatrix} \cos(\omega_k(t-\tau))I & \sin(\omega_k(t-\tau))I\\ -\sin(\omega_k(t-\tau))I & \cos(\omega_k(t-\tau))I \end{pmatrix} \begin{pmatrix} \Gamma_k\phi_T\mathbf{v}(\tau)\\ 0 \end{pmatrix} \,\mathrm{d}\tau+\mathbf{R}_k(t)\\ ={}&-\int_0^t\exp\left[ \begin{pmatrix} -\alpha I & \omega_k I\\ -\omega_k I & -\alpha I \end{pmatrix}(t-\tau)\right] \begin{pmatrix} \Gamma_k\phi_T\mathbf{v}(\tau)\\ 0 \end{pmatrix} \,\mathrm{d}\tau+\mathbf{R}_k(t). \tag*{(13)} \end{aligned}

By taking the time derivative of Eq. (13) with respect to tt, we have

dzkdt=(−αIωkI−ωkI−αI)⏟≜Jzk−(ΓkϕTv(t)0)+dRkdt−JRk(t).(14)\frac{\mathrm{d}\mathbf{z}_k}{\mathrm{d}t}= \underbrace{ \begin{pmatrix} -\alpha I & \omega_k I\\ -\omega_k I & -\alpha I \end{pmatrix}}_{\triangleq J} \mathbf{z}_k- \begin{pmatrix} \Gamma_k\phi_T\mathbf{v}(t)\\ 0 \end{pmatrix} +\frac{\mathrm{d}\mathbf{R}_k}{\mathrm{d}t}-J\mathbf{R}_k(t). \tag*{(14)}

Furthermore, we note that Rk(t)\mathbf{R}_k(t) can be modeled as a generalized Ornstein–Uhlenbeck process and dRkdt−JRk(t)\frac{\mathrm{d}\mathbf{R}_k}{\mathrm{d}t}-J\mathbf{R}_k(t) can be represented by

dRkdt−JRk(t)=ΛkW˙k,t,(15)\frac{\mathrm{d}\mathbf{R}_k}{\mathrm{d}t}-J\mathbf{R}_k(t)=\Lambda_k\dot{\mathbf{W}}_{k,t}, \tag*{(15)}

where W˙k,t\dot{\mathbf{W}}_{k,t} is the standard white noise and ΛkΛkT=−kBT(J+JT)\Lambda_k\Lambda_k^T=-k_BT(\mathbf{J}+\mathbf{J}^T). With this choice, the covariance of Rk(t)=[Rk,1,Rk,2]\mathbf{R}_k(t)=[\mathbf{R}_{k,1},\mathbf{R}_{k,2}] is given by

⟨Rk(t)Rk(τ)T⟩=kBTe−α(t−τ)(cos⁡(ωk(t−τ))Isin⁡(ωk(t−τ))I−sin⁡(ωk(t−τ))Icos⁡(ωk(t−τ))I)\left\langle\mathbf{R}_k(t)\mathbf{R}_k(\tau)^T\right\rangle =k_BT e^{-\alpha(t-\tau)} \begin{pmatrix} \cos(\omega_k(t-\tau))I & \sin(\omega_k(t-\tau))I\\ -\sin(\omega_k(t-\tau))I & \cos(\omega_k(t-\tau))I \end{pmatrix}

such that Eq. (11) remains valid. Using Eqs. (10)(12)(14), we can write the reduced model (B1) in the form of

ddt(qp⋯zk,1zk,2⋯)=(0I⋯00⋯−I0⋯ϕ(q)ΓkΓkT0⋯0⋯⋯⋯⋯⋯0−Γkϕ(q)⋯−αI−ωkI⋯00⋯ωkI−αI⋯0⋯⋯⋯⋯⋯)(∇U(q)v⋯zk,1zk,2⋯)+(00⋯ΛkW˙k,t⋯)≜K∇F(q,p,⋯ ,zk,1,zk,2,⋯ )+ΛW˙t,(16)\frac{\mathrm{d}}{\mathrm{d}t} \begin{pmatrix} \mathbf{q}\\ \mathbf{p}\\ \cdots\\ \mathbf{z}_{k,1}\\ \mathbf{z}_{k,2}\\ \cdots \end{pmatrix} = \begin{pmatrix} 0 & I & \cdots& 0 & 0 & \cdots\\ -I & 0 & \cdots& \phi(\mathbf{q})\Gamma_k\Gamma_k^T & 0 & \cdots\\ 0 & \cdots& \cdots& \cdots& \cdots& \cdots\\ 0 & -\Gamma_k\phi(\mathbf{q}) & \cdots& -\alpha I & -\omega_k I & \cdots\\ 0 & 0 & \cdots& \omega_k I & -\alpha I & \cdots\\ 0 & \cdots& \cdots& \cdots& \cdots& \cdots \end{pmatrix} \begin{pmatrix} \nabla U(\mathbf{q})\\ \mathbf{v}\\ \cdots\\ \mathbf{z}_{k,1}\\ \mathbf{z}_{k,2}\\ \cdots \end{pmatrix} + \begin{pmatrix} 0\\ 0\\ \cdots\\ \Lambda_k\dot{\mathbf{W}}_{k,t}\\ \cdots \end{pmatrix} \triangleq\mathbf{K}\nabla F(\mathbf{q},\mathbf{p},\cdots,\mathbf{z}_{k,1},\mathbf{z}_{k,2},\cdots)+\Lambda\dot{\mathbf{W}}_t, \tag*{(16)}

where K\mathbf{K} is the first matrix of the right-hand-side of Eq. (16), F(q,p,⋯ ,zk,1,zk,2,⋯ )=U(q)+12pTM−1p+12∑k=1Nω(zk,1Tzk,1+zk,2Tzk,2)F(\mathbf{q},\mathbf{p},\cdots,z_{k,1},z_{k,2},\cdots)=U(\mathbf{q})+\frac{1}{2}\mathbf{p}^{T}\mathbf{M}^{-1}\mathbf{p}+\frac{1}{2}\sum_{k=1}^{N_{\omega}}\left(z_{k,1}^{T}z_{k,1}+z_{k,2}^{T}z_{k,2}\right) is the total free energy of the extended system, and Λ=diag⁡(0,0,⋯ ,Λk,⋯ )\boldsymbol{\Lambda}=\operatorname{diag}(0,0,\cdots,\boldsymbol{\Lambda}_{k},\cdots). Using (B9), it is easy to show ΛΛT=−kBT(K+KT)\boldsymbol{\Lambda}\boldsymbol{\Lambda}^{T}=-k_{B}T(\mathbf{K}+\mathbf{K}^{T}). Therefore, the gradient system (B10) (i.e., the reduced model (B1)) has the invariant density function

ρeq(q,p,z)=exp⁡[−F(q,p,z)/kBT].\rho_{\mathrm{eq}}(\mathbf{q},\mathbf{p},\mathbf{z})=\exp\left[-F(\mathbf{q},\mathbf{p},\mathbf{z})/k_{B}T\right].

Training details of the reduced model

We represent the state encoders ϕ(q)=[ϕ1(q),⋯ ,ϕn(q)]\phi(\mathbf{q})=[\phi_{1}(\mathbf{q}),\cdots,\phi_{n}(\mathbf{q})] and Θ(t)\Theta(t) in reduced model (B1) by

ϕi(q)=HiTψ(q)Θ(t)=e−αt∑k=0NωΘ^kcos⁡(ωkt),(17)\begin{aligned} \phi_{i}(\mathbf{q})&=\mathbf{H}_{i}^{T}\psi(\mathbf{q})\\ \Theta(t)&=\mathrm{e}^{-\alpha t}\sum_{k=0}^{N_{\omega}}\widehat{\Theta}_{k}\cos(\omega_{k}t), \tag*{(17)} \end{aligned}

where ψ(q)=[ψ1(q),⋯ ,ψNb(q)]\psi(\mathbf{q})=[\psi_{1}(\mathbf{q}),\cdots,\psi_{N_{b}}(\mathbf{q})] is a set of sparse bases, H=[H1T,⋯ ,HnT]\mathbf{H}=[\mathbf{H}_{1}^{T},\cdots,\mathbf{H}_{n}^{T}] are trainable coefficients. ωk=2πk/Tc\omega_{k}=2\pi k/T_{c} is the frequency, where TcT_{c} is the time domain cut-off of the kernel and δt\delta t is the step size of the discrete samples. For the present study, ψ\psi is chosen as the uniform piecewise linear basis function defined on [2.8,4.1][2.8,4.1] with Nb=66N_{b}=66, Tc=200T_{c}=200 and Nω=2000N_{\omega}=2000.

To train the reduced model, we use the correlation functions conditional with differential initial states such that the state-dependent nature can be manifested. Specifically, we right-multiply the second equation of Eq. (7) by v(0)\mathbf{v}(0) and take the conditional expectation on q0=q∗\mathbf{q}_{0}=\mathbf{q}^{*}, which defines g(t;q∗)=⟨[p˙t+∇U(qt)]v0T∣q0=q∗⟩\mathbf{g}(t;\mathbf{q}^{*})=\langle[\dot{\mathbf{p}}_{t}+\nabla U(\mathbf{q}_{t})]\mathbf{v}_{0}^{T}\mid\mathbf{q}_{0}=\mathbf{q}^{*}\rangle. Accordingly, we define the empirical loss function

L=∑l=1Nq∑k=1Nt∥g~(tk;q(l))−g(tk;q(l))∥2,L=\sum_{l=1}^{N_{q}}\sum_{k=1}^{N_{t}}\left\|\widetilde{\mathbf{g}}(t_{k};\mathbf{q}^{(l)})-\mathbf{g}(t_{k};\mathbf{q}^{(l)})\right\|^{2},
g~(tk;q(l))=∑j=1kTr⁡[Θ(tk−tj)HCψ,ψ(tk,tj;q(l))HT]δt,\widetilde{\mathbf{g}}(t_{k};\mathbf{q}^{(l)})=\sum_{j=1}^{k}\operatorname{Tr}\left[\Theta(t_{k}-t_{j})\mathbf{H}\mathbf{C}_{\psi,\psi}(t_{k},t_{j};\mathbf{q}^{(l)})\mathbf{H}^{T}\right]\delta t,

where g~(⋅)\widetilde{\mathbf{g}}(\cdot) represents the prediction by the reduced model; ll and kk correspond to various initial states and discrete time, respectively. Cψ,ψ(t,τ;q∗)=⟨ψτvτv0TψtT∣q0=q∗⟩\mathbf{C}_{\psi,\psi}(t,\tau;\mathbf{q}^{*})=\langle\psi_{\tau}\mathbf{v}_{\tau}\mathbf{v}_{0}^{T}\psi_{t}^{T}\mid\mathbf{q}_{0}=\mathbf{q}^{*}\rangle is a three-point correlation characterizing the coupling among the bases.

Besides the conditional correlation functions, we can also introduce the loss function with respect to the overall correlation function, i.e.,

L2=∑k=1Nt∥g~2(tk)−g2(tk)∥2L_2 = \sum_{k=1}^{N_t} \left\lVert\widetilde{\mathbf{g}}_2(t_k) - \mathbf{g}_2(t_k) \right\rVert^2
g~2(tk)=∑j=1kTr⁡[Θ(tk−tj)HC‾ψ,ψ(tk,tj)HT]δt,\widetilde{\mathbf{g}}_2(t_k) = \sum_{j=1}^{k} \operatorname{Tr}\left[\boldsymbol{\Theta}(t_k-t_j)\mathbf{H}\overline{\mathbf{C}}_{\psi,\psi}(t_k,t_j)\mathbf{H}^{T}\right]\delta t,

where g2(t)=⟨[p˙t+∇U(qt)]v0T⟩\mathbf{g}_2(t)=\left\langle[\dot{\mathbf{p}}_t+\nabla U(\mathbf{q}_t)]\mathbf{v}_0^{T}\right\rangle is the overall correlation and C‾ψ,ψ(t,τ)=⟨ψτvτv0TψtT⟩\overline{\mathbf{C}}_{\psi,\psi}(t,\tau)=\left\langle\psi_{\tau}\mathbf{v}_{\tau}\mathbf{v}_0^{T}\psi_t^{T}\right\rangle. In particular, if there is scale separation between cvv(t)c_{vv}(t) and cqq(t)c_{qq}(t) (e.g., cvv(t)c_{vv}(t) decays much faster than cqq(t)c_{qq}(t); see Figure 2 and Figure 4), we may approximate C‾ψ,ψ(t,τ)\overline{\mathbf{C}}_{\psi,\psi}(t,\tau) by two-point correlations, i.e., C‾ψ,ψ(t,τ)≈⟨ψτ⊗ψtT⟩⋅⟨vτv0T⟩\overline{\mathbf{C}}_{\psi,\psi}(t,\tau)\approx\left\langle\psi_{\tau}\otimes\psi_t^{T}\right\rangle\cdot\left\langle\mathbf{v}_{\tau}\mathbf{v}_0^{T}\right\rangle.

Efficient training is achieved by using the following numerical methods to evaluate g~\widetilde{\mathbf{g}} and g~2\widetilde{\mathbf{g}}_2. Specifically, ψτψτT\psi_{\tau}\psi_{\tau}^{T} can be efficiently pre-computed with O(1)O(1) complexity by using the sparse piecewise linear basis functions. Furthermore, we can use the low-rank representation (e.g., based on the singular value decomposition) of Cψ,ψ\mathbf{C}_{\psi,\psi} and C‾ψ,ψ\overline{\mathbf{C}}_{\psi,\psi} to accelerate the matrix production HCψ,ψHT\mathbf{H}\mathbf{C}_{\psi,\psi}\mathbf{H}^{T}. In addition, the convolution on index jj can be efficiently evaluated by the Fast Fourier Transform algorithm [70].

While L2L_2 alone is insufficient to characterize the emergence of the state-dependent memory, it serves as a necessary condition and can facilitate the learning of the reduced model. In practice, we can use both loss functions to train the reduced model with Nq=65N_q=65, Nt=300N_t=300 for LL, and Nt=30000N_t=30000 for L2L_2. Specifically, the training is conducted by the Adam [77] optimization method in three stages with 2000, 6000, and 6000 steps respectively. For the first stage, we only use L2L_2 to train the model with a constant learning rate of 0.04. For each step of the following two stages, 16 initial states (i.e., q(l)\mathbf{q}^{(l)}) are randomly selected as one training batch to evaluate the total loss Lt=L+L2L_t=L+L_2. For both stages, the initial learning rate is 1×10−21\times10^{-2} and the exponential decay rate is 0.9 per 150 steps.

Simulation of the reduced model

To simulate the reduced model (B1), we follow Prop. B.1 and generate R~(t)\widetilde{\mathbf{R}}(t) on [0,T][0,T] similar to Refs. [68, 69] by

R~(t)=β−1/2∑k=02NΘ~k1/2[cos⁡(ωkt)ξk+sin⁡(ωkt)ηk],(18)\widetilde{\mathbf{R}}(t)=\beta^{-1/2}\sum_{k=0}^{2N}\widetilde{\boldsymbol{\Theta}}_k^{1/2}\left[\cos(\omega_k t)\boldsymbol{\xi}_k+\sin(\omega_k t)\boldsymbol{\eta}_k\right], \tag*{(18)}

where β−1=kBT\beta^{-1}=k_{\mathrm{B}}T, Θ~k\widetilde{\Theta}_k are the Fourier (essentially cosine) modes of Θ(∣t∣)\Theta(|t|) on [−T,T][-T,T], ξk\xi_k and ηk\eta_k are independent Gaussian random vectors, and NN is the total number of simulation step.

Specifically, for large simulation time TT, the Fourier modes Θ~k\widetilde{\Theta}_k is given by

Θ~k=∑j=1Nω∫−TTe−α∣t∣Θ^jcos⁡(ωjt)cos⁡(ωkt) dt=∑j=1Nω∫0Te−αtΘ^j(cos⁡((ωj−ωk)t)+cos⁡((ωj+ωk)t)) dt≈∑j=1Nω(αΘ^jα2+(ωj−ωk)2+αΘ^jα2+(ωj+ωk)2).\widetilde{\Theta}_k=\sum_{j=1}^{N_\omega}\int_{-T}^{T}\mathrm{e}^{-\alpha|t|}\widehat{\Theta}_j\cos(\omega_jt)\cos(\omega_kt)\,\mathrm{d}t =\sum_{j=1}^{N_\omega}\int_{0}^{T}\mathrm{e}^{-\alpha t}\widehat{\Theta}_j\left(\cos\left((\omega_j-\omega_k)t\right)+\cos\left((\omega_j+\omega_k)t\right)\right)\,\mathrm{d}t \approx\sum_{j=1}^{N_\omega}\left(\frac{\alpha\widehat{\Theta}_j}{\alpha^2+(\omega_j-\omega_k)^2}+\frac{\alpha\widehat{\Theta}_j}{\alpha^2+(\omega_j+\omega_k)^2}\right).

Therefore R~(t)\widetilde{\mathbf{R}}(t) can be generated using the Fast Fourier Transform algorithm [70] using O(Nlog⁡N)O(N\log N) complexity. Also, the convolution term ∫0tϕτTΘ(t−τ)ϕv(τ) dτ\int_{0}^{t}\phi_{\tau}^{T}\Theta(t-\tau)\phi_v(\tau)\,\mathrm{d}\tau in Eq. (7) can be computed using the fast convolution method developed in Ref. [71] with O(Nlog⁡N)O(N\log N) complexity.

Full atomistic model

We consider the full micro-scale model of benzyl bromide (see Figure 5 for a sketch of the molecule structure) in an aqueous environment. The general AMBER [78] force field is used for the benzyl bromide molecule and the partial charges of molecule atoms were set by the restrained electrostatic potential (RESP) approach [79]. The rigid TIP3P water model [80] is used for the water molecules and the bond lengths and angles were held constant through the SHAKE algorithm [81, 82]. Long-range electrostatic interactions were calculated using a Particle Mesh Ewald summation with a relative error set to be 10−410^{-4}.

The full system consists of one benzyl bromide molecule and 2400 water molecules with the periodic boundary condition imposed along each direction. The isothermal-isobaric thermostat [72] is used to equilibrate the system for 16 ns at 298K and 1 bar using a time step of 1 fs. Following the equilibration, the box size is scaled to be near 41.5×41.5×41.541.5\times41.5\times41.5 A˚3\text{\AA}^{3}. The simulation was run for a production period of 2.5 μs2.5\ \mu\mathrm{s} in a canonical ensemble with a Nosé-Hoover thermostat [73, 74].

The resolved variable qq is defined as the distance between the bromine atom and the ipso-carbon atom. The free energy is obtained from the probability density function ρ(q)\rho(q) (see the inset plot of Figure 2(a)), i.e., U(q)=−kBTln⁡ρ(q)U(q)=-k_{B}T\ln\rho(q), where ρ(q)\rho(q) is directly obtained from the full MD samples using the kernel density estimation. To verify the accuracy of the constructed U(q)U(q), we calculate the expectation of q∇U(q)q\nabla U(q) on the sample. The numerical result gives 0.996kBT0.996k_{B}T and is close to the theoretical prediction ⟨q∇U(q)⟩=∫q∇U(q)e−U(q)/kBT dq≡kBT\langle q\nabla U(q)\rangle=\int q\nabla U(q)e^{-U(q)/k_{B}T}\,\mathrm{d}q\equiv k_{B}T.

Additional numerical results

Limitations of the standard GLE near the local minima

Figure 6 shows the predictions of the conditional correlations cqv(t,q∗)c_{qv}(t,q^*) and cvv(t,q∗)c_{vv}(t,q^*) for q∗q^* representing the two local minima. Similar to the results of the saddle point as shown in Figure 3(b), the predictions of the standard GLE show apparent deviations from the full MD results due to the ignorance of the state-dependent memory nature. In contrast, the predictions of the present model with four features can accurately recover the MD predictions.

Figure 6. The conditional correlation functions cqv(t,q∗)c_{qv}(t,q^*) and cvv(t,q∗)c_{vv}(t,q^*) for the two local minima predicted by the full MD, the standard GLE, and the present model (SD-GLE) constructed using one and four spatio-features. Left: q∗=3.07q^*=3.07; Right: q∗=3.87q^*=3.87. The predictions by the standard GLE show apparent discrepancies with the full MD results.

Other forms of the reduced model

For comparison, we also consider other forms of the reduced model. In particular, we retain the encoders ϕ\phi in Eq. (7) but set Θ(t)\Theta(t) to be diagonal, i.e., we ignore the non-Markovian coupling among the different state features. The reduced model is trained using four features. Figure 7 shows the conditional correlation cvv(t;q∗)c_{vv}(t;q^*) obtained from the full MD and different reduced models. The prediction of the constructed model (labeled by “SD-GLE-Diag”) shows apparent deviations from the full MD result with incremental improvement over the stand GLE. The large discrepancy reveals the complex state-dependent nature; the non-Markovian effect can be neither approximated by ansatz like γ(q)θ(t)\gamma(q)\theta(t) as a simple generalization/re-scaling of the initial value at t=0t=0, nor represented by the coupling with the independent bath variables. Instead, the non-Markovian coupling among the various state-features retained in the present model plays a crucial for accurately modeling the heterogeneous energy dissipation arising from the unresolved intramolecular interactions and reproducing the collective dynamics.

Figure 7. The conditional correlation functions cvv(t,q∗)c_{vv}(t,q^*) for the saddle point predicted by the full MD, the standard GLE, the reduced models constructed using four state-features with diagonal Θ(t)\Theta(t) (SD-GLE-Diag) and full Θ(t)\Theta(t) (SD-GLE). The large discrepancy between the SD-GLE-Diag model and the full MD results implies the complexity of the state-dependency of the memory term, which can not be well represented by the coupling of independent bath variables. The non-Markovian interactions among the state-features are essential to capture the heterogeneous energy dissipation process.

Generalization of the present reduced model formulation

So far, we have constructed the reduced model (B1) by assuming the matrix-valued kernel Θ(t)\Theta(t) is symmetry. In fact, this form can be generalized by introducing an anti-symmetry part, i.e.,

Θ(t)=e−αt∑k=0Nω(Γk,1TΓk,1+Γk,2TΓk,2)cos⁡(ωkt)+(Γk,1TΓk,2−Γk,2TΓk,1)sin⁡(ωkt),(19)\Theta(t) = e^{-\alpha t}\sum_{k=0}^{N_\omega}\left(\Gamma_{k,1}^{T}\Gamma_{k,1}+\Gamma_{k,2}^{T}\Gamma_{k,2}\right)\cos(\omega_k t)+\left(\Gamma_{k,1}^{T}\Gamma_{k,2}-\Gamma_{k,2}^{T}\Gamma_{k,1}\right)\sin(\omega_k t), \tag*{(19)}

where Γk,1\Gamma_{k,1} and Γk,2\Gamma_{k,2} are lower-triangular matrices representing the Fourier modes of Θ(t)\Theta(t). The form is general non-symmetric except for t=0t=0 and satisfies Θ(−t)=Θ(t)T\Theta(-t)=\Theta(t)^T.

Similar to the symmetry form, we can model the fluctuation term Rt\mathbf{R}_t as a noise in the form of Rt=ϕtTR~(t)\mathbf{R}_t=\phi_t^T\widetilde{\mathbf{R}}(t), where R~(t)\widetilde{\mathbf{R}}(t) is a Gaussian random process satisfying ⟨R~(t)R~(τ)T⟩=kBTe−α(t−τ)Θ(t−τ)\langle\widetilde{\mathbf{R}}(t)\widetilde{\mathbf{R}}(\tau)^T\rangle=k_B T e^{-\alpha(t-\tau)}\Theta(t-\tau). Similar to Prop. B.1, we can show that this choice retains a consistent invariant density function.

In practice, we can generate the noise term R~(t)\widetilde{\mathbf{R}}(t) on [0,T][0,T] by

R~(t)=β−1/2∑k=02N[Θ~k,11/2cos⁡(ωkt)ξk+sin⁡(ωkt)(Q1,k1/2ξk+Q2,k1/2ηk)],\widetilde{\mathbf{R}}(t)=\beta^{-1/2}\sum_{k=0}^{2N}\left[\widetilde{\Theta}_{k,1}^{1/2}\cos(\omega_k t)\xi_k+\sin(\omega_k t)\left(Q_{1,k}^{1/2}\xi_k+Q_{2,k}^{1/2}\eta_k\right)\right],
Q1,k=Θ~k,2Θ~k,1−1Θ~k,2T,Q2,k=Θ~k,1−Θ~k,2Θ~k,1−1Θ~k,2T,Q_{1,k}=\widetilde{\Theta}_{k,2}\widetilde{\Theta}_{k,1}^{-1}\widetilde{\Theta}_{k,2}^{T},\qquad Q_{2,k}=\widetilde{\Theta}_{k,1}-\widetilde{\Theta}_{k,2}\widetilde{\Theta}_{k,1}^{-1}\widetilde{\Theta}_{k,2}^{T},

where β−1=kBT\beta^{-1}=k_{\mathrm{B}}T, Θ~k,1\widetilde{\Theta}_{k,1}, Θ~k,2\widetilde{\Theta}_{k,2} are the Fourier cosine and sine modes on [−T,T][-T,T] with Θ(−t)=Θ(t)\Theta(-t)=\Theta(t), ξk\xi_k and ηk\eta_k are independent Gaussian random vectors, and NN is the total number of simulation step. Here R~(t)\widetilde{\mathbf{R}}(t) can be also generated using the Fast Fourier Transform algorithm [70] using O(Nlog⁡N)O(N\log N) complexity. We will investigate this generalized formulation for model reduction in future studies.

References

  1. [1]B. O. Koopman, Proceedings of the National Academy of Sciences 17, 315 (1931).
  2. [2]H. Mori, Progress of Theoretical Physics 33, 423 (1965).
  3. [3]R. Zwanzig, Lectures in Theoretical Physics 3, 106 (1961).
  4. [4]R. Zwanzig, Nonequilibrium Statistical Mechanics (Oxford University Press, 2001).DOI
  5. [5]O. F. Lange and H. Grubmüller, J. Chem. Phys. 124, 214903 (2006).
  6. [6]E. Darve, J. Solomon, and A. Kia, Proc. Natl. Acad. Sci. 106, 10884 (2009).
  7. [7]M. Ceriotti, G. Bussi, and M. Parrinello, Physical review letters 102, 020601 (2009).
  8. [8]A. D. Baczewski and S. D. Bond, The Journal of chemical physics 139, 044107 (2013).
  9. [9]A. Davtyan, J. F. Dama, G. A. Voth, and H. C. Andersen, J. Chem. Phys. 142, 154104 (2015).
  10. [10]H. Lei, N. A. Baker, and X. Li, Proc. Natl. Acad. Sci. 113, 14183 (2016).
  11. [11]A. Russo, M. A. Durán-Olivencia, I. G. Kevrekidis, and S. Kalliadasssis, arXiv preprint arXiv:1903.09562 (2019).DOI
  12. [12]G. Jung, M. Hanke, and F. Schmid, Journal of Chemical Theory and Computation 13, 2481 (2017).
  13. [13]H. S. Lee, S.-H. Ahn, and E. F. Darve, The Journal of Chemical Physics 150, 174113 (2019).
  14. [14]L. Ma, X. Li, and C. Liu, Journal of Computational Physics 380, 170 (2019).
  15. [15]S. Wang, Z. Ma, and W. Pan, Soft Matter 16, 8330 (2020).
  16. [16]Y. Zhu and D. Venturi, Journal of Statistical Physics , 1217 (2020).
  17. [17]V. Klippenstein and N. F. A. van der Vegt, The Journal of Chemical Physics 154, 191102 (2021).
  18. [18]H. Vroylandt, L. Goudenège, P. Monmarché, F. Pietrucci, and B. Rotenberg, Proceedings of the National Academy of Sciences 119, e2117586119 (2022).
  19. [19]Z. She, P. Ge, and H. Lei, The Journal of Chemical Physics 158, 034102 (2023).
  20. [20]P. Xie, R. Car, and W. E, arXiv preprint arXiv:2211.06558 (2022).DOI
  21. [21]P. Hänggi, in Stochastic Dynamics, edited by L. Schimansky-Geier and T. Pöschel (Springer Berlin Heidelberg, Berlin, Heidelberg, 1997) pp. 15–22.
  22. [22]V. Klippenstein, M. Tripathy, G. Jung, F. Schmid, and N. F. van der Vegt, The Journal of Physical Chemistry B 125, 4931 (2021).
  23. [23]H. A. Posch, U. Balucani, and R. Vallauri, Physica A 123, 516 (1984).
  24. [24]J. E. Straub, M. Borkovec, and B. J. Berne, The Journal of Physical Chemistry 91, 4995 (1987).
  25. [25]J. E. Straub, B. J. Berne, and B. Roux, The Journal of Chemical Physics 93, 6804 (1990).
  26. [26]S. S. Plotkin and P. G. Wolynes, Phys. Rev. Lett. 80, 5015 (1998).
  27. [27]G. Luo, I. Andricioaei, X. S. Xie, and M. Karplus, The Journal of Physical Chemistry B 110, 9363 (2006).
  28. [28]R. B. Best and G. Hummer, Phys. Rev. Lett. 96, 228104 (2006).
  29. [29]R. B. Best and G. Hummer, Proceedings of the National Academy of Sciences 107, 1088 (2010).
  30. [30]M. Hinczewski, Y. von Hansen, J. Dzubiella, and R. R. Netz, The Journal of Chemical Physics 132, 245103 (2010).
  31. [31]R. Satija, A. Das, and D. E. Makarov, The Journal of Chemical Physics 147, 152707 (2017).
  32. [32]J. A. Morrone, J. Li, and B. J. Berne, The Journal of Physical Chemistry B 116, 378 (2012).
  33. [33]J. O. Daldrop, B. G. Kowalik, and R. R. Netz, Physical Review X 7, 041065 (2017).
  34. [34]J. M. Deutch and I. Oppenheim, The Journal of Chemical Physics 54, 3547 (1971).
  35. [35]R. Zwanzig, J. Stat. Phys. 9, 215 (1973).
  36. [36]R. Zwanzig, The Journal of Physical Chemistry 96, 3926 (1992).
  37. [37]A. Berezhkovskii and A. Szabo, The Journal of Chemical Physics 135, 074108 (2011).
  38. [38]F. Glatzel and T. Schilling, Europhysics Letters 136, 36001 (2022).
  39. [39]H. Vroylandt, Europhysics Letters 140, 62003 (2022).
  40. [40]H. Vroylandt and P. Monmarché, The Journal of Chemical Physics 156, 244105 (2022).
  41. [41]C. Ayaz, L. Scalfi, B. A. Dalton, and R. R. Netz, Physical Review E 105, 054138 (2022).
  42. [42]B. Jung and G. Jung, The Journal of Chemical Physics 159, 084110 (2023).
  43. [43]L. Lyu and H. Lei, Phys. Rev. Lett. 131, 177301 (2023).
  44. [44]R. Satija and D. E. Makarov, The Journal of Physical Chemistry B 123, 802 (2019).
  45. [45]F. Grogan, H. Lei, X. Li, and N. A. Baker, J. Comput. Phys. 418, 109633 (2020).
  46. [46]D. Singh, K. Mondal, and S. Chaudhury, The Journal of Physical Chemistry B 125, 4536 (2021).
  47. [47]C. Ayaz, L. Tepper, F. N. Brüning, J. Kappler, J. O. Daldrop, and R. R. Netz, Proceedings of the National Academy of Sciences 118, e2023856118 (2021).
  48. [48]B. A. Dalton, C. Ayaz, H. Kiefer, A. Klimek, L. Tepper, and R. R. Netz, Proceedings of the National Academy of Sciences 120, e2220068120 (2023).
  49. [49]J. E. Straub, M. Borkovec, and B. J. Berne, The Journal of Chemical Physics 89, 4833 (1988).
  50. [50]S. Singh, R. Krishnan, and G. Robinson, Chemical Physics Letters 175, 338 (1990).
  51. [51]B. Carmeli and A. Nitzan, Chemical Physics Letters 102, 517 (1983).
  52. [52]G. Tarjus and D. Kivelson, Chemical Physics 152, 153 (1991).
  53. [53]R. Krishnan, S. Singh, and G. W. Robinson, The Journal of Chemical Physics 97, 5516 (1992).
  54. [54]G. A. Voth, The Journal of Chemical Physics 97, 5908 (1992).
  55. [55]J. B. Straus, J. M. Gomez Llorente, and G. A. Voth, The Journal of Chemical Physics 98, 4082 (1993).
  56. [56]G. R. Haynes, G. A. Voth, and E. Pollak, Chemical Physics Letters 207, 309 (1993).
  57. [57]G. R. Haynes, G. A. Voth, and E. Pollak, The Journal of Chemical Physics 101, 7811 (1994).
  58. [58]P. Cossio, G. Hummer, and A. Szabo, Proceedings of the National Academy of Sciences 112, 14248 (2015).
  59. [59]G. Torrie and J. Valleau, Journal of Computational Physics 23, 187 (1977).
  60. [60]S. Kumar, J. M. Rosenberg, D. Bouzida, R. H. Swendsen, and P. A. Kollman, Journal of Computational Chemistry 13, 1011 (1992).
  61. [61]E. Darve and A. Pohorille, The Journal of Chemical Physics 115, 9169 (2001).
  62. [62]A. Laio and M. Parrinello, Proceedings of the National Academy of Sciences 99, 12562 (2002).
  63. [63]L. Rosso, P. Mináry, Z. Zhu, and M. E. Tuckerman, The Journal of Chemical Physics 116, 4389 (2002).
  64. [64]L. Maragliano and E. Vanden-Eijnden, Chemical Physics Letters 426, 168 (2006).
  65. [65]J. B. Abrams and M. E. Tuckerman, The Journal of Physical Chemistry B 112, 15742 (2008).
  66. [66]L. Maragliano and E. Vanden-Eijnden, The Journal of Chemical Physics 128, 184110 (2008).
  67. [67]H. Lei and X. Li, The Journal of Chemical Physics 154, 184108 (2021).
  68. [68]M. Berkowitz, J. Morgan, and J. A. McCammon, J. Chem. Phys. 78, 3256 (1983).
  69. [69]V. A. Ogorodnikov and S. M. Prigarin, Numerical Modelling of Random Processes and Fields: Algorithms and Applications (De Gruyter, Berlin, Boston, 1996).
  70. [70]J. W. Cooley and J. W. Tukey, Mathematics of Computation 19, 297 (1965).
  71. [71]A. Schä​dle, M. López-Fernández, and C. Lubich, SIAM Journal on Scientific Computing 28, 421 (2006).
  72. [72]G. J. Martyna, D. J. Tobias, and M. L. Klein, The Journal of Chemical Physics 101, 4177 (1994).
  73. [73]S. Nosé, Molecular Physics 52, 255 (1984).
  74. [74]W. G. Hoover, Physical Review A 31, 1695 (1985).
  75. [75]H. Kramers, Physica 7, 284 (1940).
  76. [76]W. E. and E. Vanden-Eijnden, Annual Review of Physical Chemistry 61, 391 (2010).
  77. [77]D. Kingma and J. Ba, International Conference on Learning Representations (ICLR) (2015).
  78. [78]J. Wang, R. M. Wolf, J. W. Caldwell, P. A. Kollman, and D. A. Case, Journal of Computational Chemistry 25, 1157 (2004).
  79. [79]C. I. Bayly, P. Cieplak, W. Cornell, and P. A. Kollman, The Journal of Physical Chemistry 97, 10269 (1993).
  80. [80]W. L. Jorgensen, J. Chandrasekhar, J. D. Madura, R. W. Impey, and M. L. Klein, The Journal of Chemical Physics 79, 926 (1983).
  81. [81]J.-P. Ryckaert, G. Ciccotti, and H. J. Berendsen, Journal of Computational Physics 23, 327 (1977).
  82. [82]S. Miyamoto and A. Kollman Peter, Journal of Computational Chemistry 13, 952 (2004).

Paper details

vlnc:W_754czqpp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents