Spin Dynamics from Atomistic Quantum Simulations Enrico Drigo, Marquis M. McMillan, Benjamin Pingault, Yinan Dong, F. Joseph Heremans, David D. Awschalom, Giulia Galli Abstract Optically active solid-state spin defects are promising candidates for quantum applications, however a unified theoretical framework to predict their spin dynamics at high temperatures is not yet available. Here, using Kubo linear–response theory, we derive expressions of spin-lattice and decoherence times T 1 T_{1} and T 2 T_{2} in terms of correlation functions of spin–lattice couplings. We then evaluate T 1 T_{1} and T 2 T_{2} from molecular dynamics and spin–lattice interaction time–series generated by state–of–the–art machine learning models trained on ab–initio data. Finally we measure T 1 T_{1} times for the NV center in diamond and compare experimental and theoretical results, showing excellent agreement. Introduction. —Spin defects in solids are promising platforms for advancing quantum technologies, and extending their quantum coherence to high temperatures is of great interest for quantum sensing and quantum communication applications. Recently, record coherence times at ambient conditions have been reported for a prototypical spin defect, the NV center in diamond, using isotopically purified samples 2 ; 4 ; 20 ; 14 . While it was shown that the NV spin can be optically polarized up to \mathrm{\approx 700\penalty\ K} , little is known about the physical processes that limit coherence at higher temperatures, and a unified theoretical framework to describe high–temperature spin relaxation is not yet available. Several methods have been proposed to compute spin relaxation times, including cluster correlation expansion (CCE) techniques 45 , the Redfield master equation 40 and Kubo linear response theory (LRT) 17 . CCE approaches are widely used to quantify the dissipative dynamics of a qubit in a bath of nuclear spins 46 ; 33 ; 36 ; 35 . The Redfield master equation provides a perturbative formalism for the evolution of the reduced–density matrix 40 , and has been successfully applied to spin–lattice relaxation in two–level systems at low temperatures under the approximation of one and two–phonon processes 9 ; 26 ; 22 ; 21 ; 23 ; 43 ; 44 . Kubo LRT describes the regression of conserved quantities, such as energy and magnetization, toward equilibrium and can account for the time evolution of lattice vibrations 18 ; 17 . In 1946, Bloch proposed a set of phenomenological equations for the time evolution of the spin magnetization, \mathbf{M} 8 : \dot{\mathbf{M}}(t)=\gamma\mathbf{M}(t)\times\mathbf{H}(t)-\overline{\mathbf{R}}\mathbf{M}(t), (1) where \mathbf{H} is the external magnetic field, \gamma is the gyromagnetic ratio, \overline{\mathbf{R}}=\mathrm{diag}\left\{\frac{1}{T_{2}},\frac{1}{T_{2}},\frac{1}{T_{1}}\right\} , and T_{2} and T_{1} are the transverse (decoherence) and longitudinal (spin–lattice) relaxation times, respectively. At finite temperatures, the irreversible energy exchange between spins and lattice causes longitudinal relaxation 8 ; 12 ; 17 . In a similar fashion, the coupling between lattice vibrations and magnetic degrees of freedom affects the precession of \mathbf{M} , leading to decoherence 12 ; 17 . In this Letter, within Kubo LRT, we identify the external perturbations responsible for spin relaxation and derive general expressions for spin–lattice and decoherence times T_{1} and T_{2} as a function of temperature. We then compute spin relaxation times of the NV center in diamond, in the absence of magnetic noise, across a broad range of thermodynamic conditions using machine–learning (ML) molecular dynamics (MD) simulations and neural–network (NN) models of spin–lattice interactions. The effect of magnetic noise on T_{2} is included at T=0 using results obtained with the CCE approach. We also measure T_{1} as a function of temperature for NV centers in diamond and compare our numerical and experimental results, finding excellent agreement. Theoretical framework. —We consider a spin–lattice system characterized by its total energy (E) and magnetization ( \mathbf{M} ), and subject to an external perturbation applied adiabatically. In the presence of an external field, E and \mathbf{M} differ from their respective equilibrium values 18 ; 17 , and upon removal of the driving field, after reaching an out-of-equilibrium steady state, E and \mathbf{M} attain their equilibrium values 17 ; 13 ; 12 . In the out–of–equilibrium steady state an energy imbalance is present between spins and lattice 13 ; 17 , and spin–lattice and decoherence relaxation processes characterize the regression of energy and transverse magnetization toward equilibrium, respectively 13 ; 17 ; 12 . Spin and lattice degrees of freedom are described by the Hamiltonian \mathcal{H} , \displaystyle\mathcal{H}=\mathcal{H}_{\mathrm{S}}^{0}+\mathcal{H}_{\mathrm{L}}+\mathcal{H}_{\mathrm{SL}}, (2) where \mathcal{H}_{\mathrm{S}}^{0} is the frozen–lattice spin Hamiltonian, \mathcal{H}_{\mathrm{L}} is the lattice contribution, \commutator{\hh_\mathrm{S}^0}{\hh_\mathrm{L}}=0 and \mathcal{H}_{\mathrm{SL}} is the interaction between magnetic and vibrational degrees of freedom. Introduced in Refs. 38 ; 27 , spin Hamiltonians, \mathcal{H}_{\mathrm{S}} , characterize perturbatively the effect of nuclear and electronic spins on the degenerate Born–Oppenheimer ground state. In general, \mathcal{H}_{\mathrm{S}} is a parametric function of the atomic positions and therefore has an implicit dependence on the lattice degrees of freedom. Following Refs. 19 ; 16 , we define the frozen–lattice spin Hamiltonian \mathcal{H}_{\mathrm{S}}^{0}\equiv\expectationvalue{\hh_\mathrm{S}}_{L} and spin–lattice coupling \mathcal{H}_{\mathrm{SL}}\equiv\mathcal{H}_{\mathrm{S}}-\expectationvalue{\hh_\mathrm{S}}_{L} . \expectationvalue{\cdot}_{L} is the expectation value computed over the lattice degrees of freedom. In the limit of high temperatures and at second order in \mathcal{H}_{\mathrm{SL}} , the Kubo LRT formulation of spin relaxation yields 17 : \displaystyle\frac{1}{T_{1}}=\sum_{\omega_{\delta}\neq 0}\omega_{\delta}^{2}\int_{0}^{\infty}\differential{t}\,e^{i\omega_{\delta}t}\frac{\expectationvalue{\hh^{(\delta)}_\mathrm{SL}(t)\hh^{(-\delta)}_\mathrm{SL}}}{\expectationvalue{{\hh_\mathrm{S}}^2}}, (3) \displaystyle\frac{1}{T_{2}}=\sum_{\omega_{\delta}}\int_{0}^{\infty}\differential{t}\,e^{i\omega_{\delta}t}\frac{\expectationvalue{\comm{\bm{\mathcal{M}}^\perp}{\hh^{(\delta)}_\mathrm{SL}(t)}\cdot\comm{\bm{\mathcal{M}}^\perp}{\hh^{(-\delta)}_\mathrm{SL}}}}{\hbar^{2}\expectationvalue{\abs{\bm{\mathcal{M}}^\perp}^2}}. (4) where we note that according to Luttinger’s theory of thermal transport, the perturbation inducing the out–of–equilibrium energy distribution is proportional to the spin Hamiltonian 24 ; 17 . In Eqs.( 3 )–( 4 ), \bm{\mathcal{M}}^{\perp} is the transverse magnetization operator, \expectationvalue{\cdot} is the expectation value computed over \mathcal{H}_{0}=\mathcal{H}_{\mathrm{S}}^{0}+\mathcal{H}_{\mathrm{L}} , \mathcal{H}^{(\delta)}_{\mathrm{SL}} is the projection of \mathcal{H}_{\mathrm{SL}} between two eigenstates of \mathcal{H}_{\mathrm{S}}^{0} , (n,m)\equiv\delta , and \omega_{\delta}=\frac{E_{n}-E_{m}}{\hbar} . In the interaction picture relative to \mathcal{H}_{0} , the evolution of \mathcal{H}_{\mathrm{SL}} is separable between spins and lattice: \mathcal{H}_{\mathrm{SL}}(t)=\sum_{\delta}e^{i\omega_{\delta}t}\mathcal{H}^{(\delta)}_{\mathrm{SL}}(t) where \mathcal{H}_{\mathrm{SL}}^{(\delta)}(t)\equiv e^{i\mathcal{H}_{\mathrm{L}}t/\hbar}\mathcal{H}_{\mathrm{SL}}^{(\delta)}e^{-i\mathcal{H}_{\mathrm{L}}t/\hbar} . The time evolution of the lattice can be sampled via lattice dynamics or MD approaches depending on the temperature regime and level of accuracy of interest. At low temperatures, phonon quasi particles accurately describe the crystal. At high temperatures, the dynamics of lattice vibrations can be estimated by integrating the classical equations of motion of the nuclei. In the following, we present a molecular dynamics approach based on ML force fields and NN models of the zero–field–splitting (ZFS) tensor to compute spin relaxation times at finite temperature, in the absence of a magnetic bath, with ab–initio accuracy. We then carry out measurements on the isotopically purified NV center in diamond and show that our theoretical and experimental results are in excellent agreement. Results. —For the NV center in diamond, the major contribution to the spin Hamiltonian comes from the ZFS and can be written as: \mathcal{H}_{\mathrm{ZFS}}=\bm{\mathcal{S}}\overline{\bm{\mathcal{D}}}\bm{\mathcal{S}}^{\mathrm{T}} , where \bm{\mathcal{S}} is the spin vector operator and \overline{\bm{\mathcal{D}}} is the ZFS tensor 38 ; 27 ; 30 ; 31 ; 32 ; 39 . In this system, spin–orbit coupling contributions to ZFS are negligible compared to spin-spin ones 30 ; 31 ; 32 ; 39 . The dipolar interaction leads to the splitting of the NV’s triplet ground state, {}^{3}A_{2} , into the \left\{\ket{0},\ket{+},\ket{-}\right\} sub–levels. The first–principle (FP) evaluation of the ZFS tensors and of MD simulation trajectories has a high computational cost, as reliable estimates of spin relaxation times require long trajectories (on the order of ns) and large supercells with hundreds of atoms. Hence, straightforward FP approaches are impracticable, and here we rely on ML methods to accelerate both MD simulations and ZFS calculations, achieving ab–initio accuracy at a fraction of the computational cost. To evaluate Eqs. ( 3 )–( 4 ), we train a MACE ML interatomic potential (MLIP) to approximate the FP potential energy surface, and a NN model to compute the ZFS tensors 6 . The MACE MLIP is trained to reproduce ab–initio total energies and atomic forces computed at the spin–polarized DFT/PBE level of theory 10 ; 37 . First, we run \mathrm{2} to \mathrm{10\penalty\ ps} Born–Oppenheimer (BO) MD simulations of one NV center in diamond cells of 64 and 512 atoms in the NVE and NVT ensembles in the \mathrm{1000} – \mathrm{2500\penalty\ K} temperature range 10 . Further, we apply an active learning scheme to optimally add structures to the training dataset using an uncertainty estimation approach 7 ; 1 . All spin Hamiltonian calculations are performed using the pyZFS code and spin polarized wave–functions at the PBE level of theory 10 ; 25 ; 37 . The ZFS NN model is a message passing equivariant graph NN based on the MACE architecture 6 ; 1 . Equivariance is imposed exactly via contractions with Clebsh–Gordan coefficients matrices, ensuring that the final output transforms under rotations as a spherical harmonics of order 2 6 . Since the ZFS tensor depends only on the local environment of the defect, we enforce a local attention criterion using the Frobenius norm of the predicted atomic tensors \overline{\bm{\mathcal{T}}}(\bm{R}) . We define the center as the atom with the highest attention. The atoms within a given cutoff from the center are labeled active . The output of the NN is the sum of the contributions of the atoms in the active region, \mathcal{A} . The remaining atomic tensors are introduced in the loss function, \mathcal{L} , as a penalty. The loss function, over the training set, \mathcal{D} , is: \mathcal{L}=\sum_{i\in\mathcal{D}}\norm{\overline{\mathbf{D}}_i-\sum_{j\in\mathcal{A}}\overline{\bm{\mathcal{T}}}(\bm{R}_j^i)}^{2}+\norm{\sum_{j\notin\mathcal{A}}\overline{\bm{\mathcal{T}}}(\bm{R}_j^i)}^{2} . Thanks to the attention criterion, the ZFS NN prediction is local. Indeed, the ZFS tensor depends only on the concentration of defects and its elements converge to finite values in the dilute limit. The extension of the ZFS NN model to multiple spin defects is currently under investigation. Figure 1: Left panel: comparison of the vibrational density of states (VDOS) of the NV center in diamond computed from molecular dynamics (MD) simulations with machine learned (ML) potentials in the dilute limit, and of crystalline diamond computed via density functional perturbation theory (DFPT) 5 . Right panel: comparison of the phonon dispersion of the NV center in diamond computed from MD in the dilute limit and of crystalline diamond computed via DFPT 5 . The ML MD calculations are performed using a \mathrm{2\penalty\ ns} NVE trajectory of one NV center in a diamond cell of 5832 atoms at 1000\penalty\ \mathrm{K} . The MACE MLIP is benchmarked in the dilute limit against the vibrational density of states (VDOS) and phonon dispersion of bulk diamond (lattice constant = 3.567\penalty\ \text{\AA} ), obtained from density functional perturbation theory (DFPT) 5 . We compute the ML VDOS and phonon dispersion from a 2\penalty\ \mathrm{ns} ML MD trajectory at \mathrm{1000\penalty\ K} of one NV center in a diamond cell of 5832 atoms harvesting configurations every timestep. In the quasi–harmonic approximation, the phonon dispersion is estimated from the poles of the Fourier transform of the time correlation function of the projection of the atomic displacements over a plane wave with wave–vector in the \Gamma – \mathrm{L} direction of the crystal. The comparison between DFPT and ML MD results in the dilute limit, presented in Fig. 1 , shows excellent agreement. We also investigated the temperature dependence of the ZFS axial coefficient, D , from 500\penalty\ \mathrm{ps} –long NVE trajectories of one NV center in a diamond cell of 512 atoms, and compared it with experimental results 3 ; 20 . The D coefficients are computed using the ZFS NN model on configurations sampled every \mathrm{10\penalty\ fs} . Our calculations show a relative shift of \approx\mathrm{150\penalty\ MHz} between the values of D obtained from ML MD calculations and experiments. This difference does not reflect the inaccuracy of the ZFS NN model but rather the limitations of the PBE functional used in computing the magnitudes of the elements of the ZFS tensor. Importantly, the evaluation of \frac{\partial D}{\partial T} at room temperature from ML MD and ZFS NN, \approx\mathrm{-100\penalty\ \frac{kHz}{K}} , is in good agreement with experimental findings, \mathrm{-76(1)\penalty\ \frac{kHz}{K}} 3 and \mathrm{-80\penalty\ \frac{kHz}{K}} 42 . After these successful benchmarks, we used the trained MACE MLIP and ZFS NN to carry out ns–long ML MD simulations and to evaluate the corresponding ZFS time–series of one NV center in diamond in a wide temperature interval spanning 400\penalty\ to 1000\penalty\ \mathrm{K} . Initially, the system is equilibrated in the NVT ensemble for 100\penalty\ \mathrm{ps} and then 16\penalty\ \mathrm{ns} –long trajectories are collected in the NVE ensemble. From the ZFS time-series obtained with the MACE MLIP and the ZFS NN model, we compute spin–lattice and decoherence relaxation times at different temperatures. Figure 2: Convergence analysis of T_{1} as a function of the number of atoms in the simulation cell and segment size used in the block analysis. All calculations are performed using \mathrm{16\penalty\ ns} NVE molecular dynamics simulations with machine learned potentials at \mathrm{500\penalty\ K} . To analyze the effect of cell size and simulation times on our results, we focused on trajectories at \mathrm{500\penalty\ K} where acoustic vibrational modes are the most populated and large supercells are required to describe low frequency modes. Error bars are estimated via block–analysis. We considered supercells up to 10648 atoms and divided the trajectories in segments of \mathrm{0.5\penalty\ ns} , \mathrm{1\penalty\ ns} and \mathrm{2\penalty\ ns} . The complete convergence analysis presented in Fig. 2 indicates that at 500\penalty\ \mathrm{K} there are significant finite size effects for diamond cells with less than 5000 atoms and that the minimal block length is 1\penalty\ \mathrm{ns} . Therefore, thereinafter all numerical results are computed using simulation cells of at least 10648 atoms and \mathrm{1\penalty\ ns} -long blocks. It is informative to decompose Eq.s ( 3 )–( 4 ) in secular and non–secular contributions. The secular terms, associated to \omega_{\delta}=0 , describe eigenvalue fluctuations. The non secular ones are responsible for transitions between sub–levels of the triplet state of the NV center. Furthermore, Eqs. ( 3 )–( 4 ) can be expanded in the frequency domain, \omega , via Fourier transforms: 1/{T_{1,2}}=\sum_{\omega_{\delta}}\lim_{\omega\to\omega_{\delta}}{T_{1,2}}^{-1}_{(\delta)}(\omega) . Our analysis highlights that spin–lattice relaxation depends only on non–secular contributions whereas T_{2} is influenced by both 34 . The spin–lattice relaxation obtained in our formulation, being a pure non–secular process, is consistent with that derived within the Redfield formalism where T_{1} is interpreted as the characteristic time associated to the decay of the diagonal elements of the density matrix of a two level system. The secular part of T_{2} is usually referred to as a pure dephasing time, T^{-1}_{\Phi}=\sum_{\omega_{\delta}=0}\lim_{\omega\to\omega_{\delta}}{T_{2}}^{-1}_{(\delta)}(\omega) . The non–secular term arises from population exchanges, {T^{\prime}_{1}}^{-1}=\sum_{\omega_{\delta}\neq 0}\lim_{\omega\to\omega_{\delta}}{T_{2}}^{-1}_{(\delta)}(\omega) , which contribute to decoherence, 1/T_{2}=T^{-1}_{\Phi}+{T^{\prime}_{1}}^{-1} . Figure 3: Frequency and transition decomposition of Eq. ( 3 ) between \ket{+} – \ket{0} and \ket{-} – \ket{0} (upper and upper middle panels), of the pure dephasing contribution to T_{2} , T_{\Phi} (lower middle panel) and of T_{1}^{\prime} (lower middle panel). All calculations are performed using a \mathrm{16\penalty\ ns} NVE molecular dynamics simulation with machine learned potentials of one NV center in a diamond cell of 10648 atoms at 500\penalty\ \mathrm{K} . In Fig. 3 , we show the sub–levels and frequency decomposition of spin relaxation times computed at \mathrm{500\penalty\ K} for one NV center in diamond. As expected, the contributions of the transitions \ket{0} – \ket{\pm} to spin–lattice relaxation are of a similar magnitude, whereas pure dephasing , T_{\Phi}^{-1} , is about one order of magnitude larger than the population term, {T^{\prime}_{1}}^{-1} , in T_{2}^{-1} . Interestingly, the Fourier decompositions of Eq.s ( 3 )–( 4 ) computed from ML MD simulations reveal the influence of low frequency lattice vibrations on the dynamics of the ZFS tensor. Indeed, in Fig. 3 we do not observe vibrational peaks after the optical resonance of crystalline diamond at \mathrm{40\penalty\ THz} . Finally, we discuss the dependence of spin relaxation times on temperature, as obtained from ML MD simulations of one NV center in diamond, and from experiments conducted on purified diamond samples 1 . The results are displayed in Fig. 4 , where we also present additional experimental data from the literature 4 ; 15 ; 42 . Figure 4: Spin-lattice relaxation, T_{1} , (upper panel) and decoherence time, T_{2} , (lower panel) of the NV center in diamond as a function of temperature. T_{1} and T_{2} are evaluated using Eqs. ( 3 )–( 4 ) from \mathrm{16\penalty\ ns} NVE molecular dynamics simulations of one NV center in a diamond cell of 10648 atoms. We compare the numerical results of T_{1} (blue circles) with our experimental results (purple stars), and with those of Ref. 4 (red star), Ref. 15 (orange stars) and Ref. 42 (green star). We omit the result at \mathrm{1000\penalty\ K} , presented in Ref. 20 , as measurements were conducted on nanodiamonds, where relaxation mechanisms may substantially differ from those in bulk diamond. The purple solid line is the low temperature fit of T_{1}^{-1}\approx aT^{5} for T\leq\mathrm{400\penalty\ K} ; the dotted blue line is the high temperature fit of T_{1}^{-1}\approx bT^{2} for T\geq\mathrm{550\penalty\ K} . We compare computed T_{2} results (blue circles) with the measurements of Ref. 4 , and show the T_{\phi} contribution (cyan circles) and T_{1}^{\prime} (brown circles). As expected, we find that spin relaxation times present a significant temperature dependence, as they drastically decrease as temperature increases. In our MD calculations, decoherence is caused by the interaction of the central spin with the lattice in the absence of magnetic noise and it is dominated by pure dephasing at all temperatures. The effect of temperature on T_{2} at \mathrm{500\penalty\ K} is T_{2}=\mathrm{414\pm 87\mu s} . We include the effect of the magnetic noise on T_{2} using the Matthiessen’s rule and computing the magnetic noise contribution with the Cluster Correlation expansion (CCE) method. Compared to the CCE results for the NV center in diamond at natural \mathrm{[^{13}C]} abundance and \mathrm{0\penalty\ K} 29 , we find that temperature effects at \mathrm{500\penalty\ K} reduce T_{2} by a factor of 4. The T_{1} value computed from MD simulations at \mathrm{600\penalty\ K} , \mathrm{0.86\pm 0.23\penalty\ ms} , is in fair agreement with measurements conducted at the same temperature, \mathrm{340\pm 50\mu s} 42 . Instead, our computed T_{1} at \mathrm{500\penalty\ K} , \mathrm{1.11\pm 0.28\penalty\ ms} , is in quantitative agreement with previous experimental results, \mathrm{0.94\pm 0.11\penalty\ ms} at \mathrm{475\penalty\ K} 15 . When considering 100\%\mathrm{[^{13}C]} in diamond at \mathrm{500\penalty\ K} , we found a moderate increase of T_{1} to \mathrm{1.26\pm 0.27\penalty\ ms} , reflecting solely the effect of mass change in the sample. The experimental measurements performed in this work spanning \mathrm{300\penalty\ K} to \mathrm{400\penalty\ K} allow us to unify the analysis of the spin lattice relaxation time as a function of temperature. Notably, at \mathrm{400\penalty\ K} our numerical result for T_{1} , \mathrm{1.41\pm 0.29\penalty\ ms} , is in excellent agreement with our experiments \mathrm{1.75\pm 0.22\penalty\ ms} , and previous measurements, \mathrm{1.67\pm 0.20\penalty\ ms} 15 . We note that prior lattice dynamics calculations considering only two–phonons processes, while accurate at low temperatures, show only a fair agreement with experiments around \mathrm{400\penalty\ K} 28 ; 9 . Under the assumption of Raman processes by spin-one-phonon interaction in second order, T_{1}^{-1} at low temperatures follows the well known T^{5} temperature behavior and at high temperatures a T^{2} scaling law 41 . In the case of diamond, whose Debye temperature is \approx\mathrm{2000\penalty\ K} , the crossover temperature from the T^{5} to T^{2} scaling is found approximately at \approx\mathrm{500\penalty\ K} . In Fig. 4 we show that the experimental measurements at T\leq\mathrm{400\penalty\ K} are accurately fitted by T_{1}^{-1}\approx aT^{5} (solid purple line) and those at T\geq\mathrm{550K} by T_{1}^{-1}\approx bT^{2} (dotted blue line). We note that the experimental result at \mathrm{600\penalty\ K} reported by Ref. 42 does not follow the expected T^{2} scaling, when considered together with the other data. A possible explanation for the disagreement is the poor contrast and performance drop at these high temperatures. Discussion. —In summary, we derived spin–lattice and decoherence relaxation times within Kubo linear response theory, describing both processes as the regression of two conserved quantities towards equilibrium, spin energy and transverse magnetization. By interpreting these phenomena as responses to external driving fields, the use of Kubo LRT unifies the description of spin-lattice relaxation and decoherence in a single theoretical framework, as shown by Eqs. ( 3 )–( 4 ), where we expressed T_{1} and the lattice contribution to T_{2} in terms of equilibrium time-correlation functions of spin–lattice couplings. We emphasize that the study of spin relaxation had so far been limited to low temperatures, using the Redfield ME and CCE methods 40 ; 22 ; 26 ; 34 ; 33 ; 35 ; 36 ; 29 ; 28 ; 43 ; 44 ; 9 ; 21 ; 23 . However, CCE techniques only account for the decoherence induced by the nuclear spins bath in the frozen lattice approximation 34 ; 33 ; 35 ; 36 ; 29 . Further, the Redfield ME relies on specific assumptions about one and two phonons processes 28 ; 22 ; 26 ; 40 ; 21 ; 23 ; 43 ; 44 ; 9 , and corrections of the Redfield ME at higher temperatures, where the harmonic approximation is no longer valid, are computationally intensive 11 ; 21 . Here, for the first time, we studied the coherence properties of the NV center in diamond at high temperatures combining Kubo LRT and MD simulations. Crucially, using ML methods to accelerate both MD simulations and the calculations of ZFS tensors, we could perform ns–long simulations for cells with \approx 10000 atoms. Our framework takes into account all phonon processes and anharmonicity. Notably, in the phonon quasi–particle limit of a two level system, Eq. ( 3 ) is equivalent to the second order Redfield ME. Our results show, as expected, that T_{1} depends only on population exchange within the sub–levels of the triplet state, while T_{2} is dominated by secular pure dephasing. Our numerical predictions at \mathrm{500\penalty\ K} and \mathrm{400\penalty\ K} are in excellent agreement with our own experiments and previous measurements 15 . Importantly, the framework presented here, based on Kubo LRT, is not restricted to ZFS Hamiltonians and can be extended to any system, spin-lattice interaction, and to any distribution of isotopic disorder. Further it can be readily applied to any triplet color centers in the solid state or to molecular qubits. In addition, thanks to the comprehensive use of NN models, our MD approach can approximate any ab–initio level of theory via active learning techniques. Acknowledgements. Acknowledgments. —We gratefully acknowledge fruitful discussions with David M. Toyli, Stefano Villani, Swarnabha Chattaraj, Lien T. Le, Jonah Nagura, Alfredo Fiorentino and Stefano Baroni. The theoretical, experimental and computational work was supported by the Midwest Integrated Center for Computational Materials (MICCoM) as part of the Computational Materials Sciences Program funded by the U.S. Department of Energy (E. D., F.J.H and G. G.). This research used resources of the National Energy Research Scientific Computing Center (NERSC), a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. ERCAP0036175 and resources of the University of Chicago Research Computing Center (RCC). Additional support for experimental validation was provided by Q-NEXT (M.M.M.); the Quantum Fellowship from the University of Chicago and the AFOSR MURI under award No. FA9550-23-1-0330 (Y.D.); and the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division through Argonne National Laboratory (B.P. and D.D.A.) under Contract No. DE-AC02-06CH11357. Data availability —The code, inputs and data that support the findings of this article are openly available at [DOIs]. End Matter Appendix A: Derivation of Eq. ( 4 ) and Eq. ( 3 ) —We consider a system described by an Hamiltonian \mathcal{H} , \displaystyle\mathcal{H}=\mathcal{H}_{0}+\mathcal{H}^{\prime}, (A1) where \mathcal{H}_{0} is the reference Hamiltonian and \mathcal{H}^{\prime} is an interaction term, \commutator{\hh_0}{\hh^\prime}\neq 0 . We study the evolution of an observable \mathcal{A} , such that \commutator{\ha}{\hh_0}=0 , under an external perturbation that couples with \mathcal{A} , \mathcal{H}_{ext}=\mathcal{A}\lambda . Within linear order in the external perturbation, and in the limit of high temperatures, we can write 17 ; 18 : \displaystyle\dot{A}\equiv\expectationvalue{\dot{\ha}} \displaystyle=\beta\int_{0}^{\infty}\left\langle\dot{\mathcal{A}}(t)\dot{\mathcal{A}}\right\rangle\differential{t}\lambda (A2) where \beta={1}\over{KT} . In the interaction picture with respect to \mathcal{H}_{0} the correlation function in Eq. ( A2 ) is: \displaystyle\left\langle\dot{\mathcal{A}}(t)\dot{\mathcal{A}}\right\rangle= \displaystyle\left\langle U^{\dagger}_{I}(t)\left(\frac{i}{\hbar}\commutator{\hh^\prime(t)}{\ha_0(t)}\right)U_{I}(t)\left(\frac{i}{\hbar}\commutator{\mathcal{H}^\prime}{\ha}\right)\right\rangle (A3) where on the left-hand side of Eq. ( A3 ) the operators are in the Heisenberg picture, while on the right-hand side of Eq. ( A3 ) operators are represented in the interaction picture: {\mathcal{A}_{0}} denotes the operator \mathcal{A} in the interaction picture and U_{I}(t) is the interaction-picture time-evolution operator, \displaystyle U_{I}(t)=1-\frac{i}{\hbar}\int_{0}^{t}\mathcal{H}^{\prime}(t^{\prime})U_{I}(t^{\prime})\differential{t^\prime}. (A4) Within second order in the interaction \mathcal{H}^{\prime} , we write: \displaystyle\left\langle\dot{\mathcal{A}}(t)\dot{\mathcal{A}}\right\rangle=\left(\frac{i}{\hbar}\right)^{2}\left\langle\commutator{\mathcal{H}^\prime(t)}{\ha_0(t)}\commutator{\mathcal{H}^\prime}{\ha}\right\rangle+\mathcal{O}({\mathcal{H}^{\prime}}^{3}). (A5) Therefore, in the limit of high temperatures, at first order in the external perturbation and at second order in the interaction, Kubo LRT yields 17 \displaystyle\dot{A} \displaystyle=\frac{\beta}{\hbar^{2}}\int_{0}^{\infty}\left\langle\commutator{{\ha_0}(t)}{\mathcal{H}^\prime(t)}\commutator{\mathcal{H}^\prime}{\ha}\right\rangle\differential{t}\lambda. (A6) which reduces to Eq.3 and Eq.4 in the main text when we substitute \mathcal{A}\equiv\mathcal{H}_{\mathrm{S}}^{0} and \mathcal{A}\equiv\bm{\mathcal{M}}^{\perp} , respectively. Appendix B: Temperature dependence of Raman scattering by spin–one–phonon interaction and spin–two–phonon interaction —Two phonon processes are the dominant decoherence mechanisms in the NV center in diamond 4 ; 15 . Specifically, for the NV center in diamond, the experimental T^{-5} behavior of T_{1} at low temperatures has been attributed to Raman processes by spin–one–phonon interaction at the second order 4 ; 15 ; 41 ; 20 . In order to study Raman processes by spin–one–phonon interaction, we expand the spin–lattice interaction at first order in the displacement Q_{\mu} as \displaystyle\mathcal{H}_{\mathrm{SL}}^{1}\equiv\sum_{\mu}\left.\frac{\partial\mathcal{H}_{\mathrm{S}}}{\partial Q_{\mu}}\right|_{0}\left(\mathcal{Q}_{\mu}+\mathcal{Q}_{\mu}^{\dagger}\right) (A7) where \mathcal{Q}_{\mu} is the creation operator of a phonon with energy \hbar\omega_{\mu} . The probability for a spin state a to be excited in a virtual state c by absorption of a phonon, and subsequently for the level c to decay to the desired state b by emission of a phonon, P_{ab} , is 41 \displaystyle P_{ab}=\frac{2\pi}{\hbar}\left|\frac{\bra{b, n_{\bm{k}}-1, n_{\bm{k}^\prime}+1}\mathcal{H}_{\mathrm{SL}}^{1}\ket{c,n_{\bm{k}}-1, n_{\bm{k}^\prime}}\bra{c,n_{\bm{k}}-1, n_{\bm{k}^\prime}}\mathcal{H}_{\mathrm{SL}}^{1}\ket{a, n_{\bm{k}}, n_{\bm{k}^\prime}}}{E_{a}-E_{c}-\omega_{\bm{k}}}\right|^{2}\times\delta(E_{a}-E_{b}-\omega_{\bm{k}}+\omega_{\bm{k}^{\prime}}). (A8) In the Debye approximation and in the limit of E_{c}-E_{a}\ll\omega_{\bm{k}} and E_{b}-E_{a}\ll\omega_{\bm{k}} , Eq. ( A8 ) reduces to 41 \displaystyle P_{ab}\propto\left(\frac{T}{\Theta_{D}}\right)^{5}\int_{0}^{\Theta_{D}/T}\frac{e^{x}x^{4}}{\left(e^{x}-1\right)^{2}}\differential{x}, (A9) where \Theta_{D} is the Debye temperature of the material. In Fig. 6 we plot the right-hand side of Eq. ( A9 ) as a function of temperature and show that for low temperatures we recover the expected T^{5} scaling law 41 . At high temperatures, Eq. ( A9 ) shows T_{1}^{-1}\propto T^{2} 41 . In diamond, \Theta_{D}\approx\mathrm{2000\penalty\ K} , and the crossover temperature of Eq. ( A9 ) from T^{5} to T^{2} scaling is found at \approx\mathrm{500\penalty\ K} . Ref. 28 and Ref. 9 have investigated the role of Raman spin–two–phonon scattering on T_{1} . The Spin Hamiltonian is expanded at second order in the displacements: \displaystyle\mathcal{H}_{\mathrm{SL}}^{2}\equiv\sum_{\mu,\nu}\left.\frac{\partial^{2}\mathcal{H}_{\mathrm{S}}}{\partial Q_{\mu}\partial Q_{\nu}}\right|_{0}\left(\mathcal{Q}_{\mu}+\mathcal{Q}_{\mu}^{\dagger}\right)\left(\mathcal{Q}_{\nu}+\mathcal{Q}_{\nu}^{\dagger}\right) (A10) In the Debye approximation and in the limit of E_{b}-E_{a}\ll\omega_{\bm{k}} , the probability of scattering from the state a to b via emission and absorption of a phonon is: \displaystyle P_{ab}\propto\left(\frac{T}{\Theta_{D}}\right)^{7}\int_{0}^{\Theta_{D}/T}\frac{e^{x}x^{6}}{\left(e^{x}-1\right)^{2}}\differential{x}. (A11) Fig. 6 displays the right-hand side of Eq. ( A11 ) as a function of temperature and at low temperatures we recover the expected T^{7} scaling law 41 . As for the Raman spin–one–phonon scattering in second order, at high temperatures, Eq. ( A11 ) shows T_{1}^{-1}\propto T^{2} 41 . Figure 5: Eq. ( A9 ) as a function of temperature for \Theta_{D}=\mathrm{1000\penalty\ K},\mathrm{1500\penalty\ K},\mathrm{2000\penalty\ K} . The solid blue line represents the analytical result, the gray dashed and dash–dotted lines show the T^{2} and T^{5} scaling laws, respectively. Figure 6: Eq. ( A11 ) as a function of temperature for \Theta_{D}=\mathrm{1000\penalty\ K},\mathrm{1500\penalty\ K},\mathrm{2000\penalty\ K} . The solid blue line represents the analytical result, the gray dashed and dash–dotted lines show the T^{2} and T^{7} scaling laws, respectively.