Atom-centered Density Matrix Propagation (ADMP): Theory and Application to protonated water clusters and water/vacuum interfaces Srinivasan S. Iyengar Department of Chemistry, Indiana University This presentation is meant to be a quick outline of ADMP. You should read the related papers to get more complete understanding Brief outline of ab initio molecular dynamics Atom-centered Density Matrix Propagation (ADMP) Results: • Novel findings for protonated water clusters • Preliminary results for ion-transport through biological channels Nut-n-bolts issues Molecular dynamics in Chemistry Molecular motion and structure determine properties: • Spectroscopic properties • Predicting Molecular Reactivity Computationally molecular dynamics simulates molecular motions: • determine properties from correlation functions • To Simulate molecular motions: – Need Energy of conformation – Forces to move nuclei: Simulate nuclear motion Methods for molecular dynamics on a single potential surface Parameterized force fields (e.g. AMBER, CHARMM) • • • Energy, forces: parameters obtained from experiment Molecules moved: Newton’s laws Works for large systems – But hard to parameterize bond-breaking/formation (chemical reactions) – Issues with polarization/charge transfer/dynamical effects Born-Oppenheimer (BO) Dynamics • • • Solve electronic Schrödinger eqn within some approximation for each nuclear structure Nuclei are propagated using gradients (forces) Works for bond-breaking but computationally expensive Large reactive, polarizable systems: We need something like BO, but less expensive. Atom-centered Density Matrix Propagation (ADMP) : An Extended Lagrangian approach Circumvent Computational Bottleneck of BO repeated SCF for electronic SE electronic structure, not converged, but propagated “Simultaneous” propagation of electronic structure with nuclei: an adjustment of timescales Avoid Atom-centered Density Matrix Propagation (ADMP) Construct a classical phase-space {{R,V,M},{P,W,m}} The Lagrangian (= Kinetic minus Potential energy) 1 1 T L Tr V MV Tr μ1/4Wμ1/4 2 2 2 Nuclear KE “Fictitious” KE of P Energy functional E(R, P) Tr Λ P 2 P Lagrangian Constraint for N-representability of P: Idempotency and Particle number P : represented using atom-centered gaussian basis sets Euler-Lagrange equations of motion Equations of motion for density matrix and nuclei “Fictitious” mass of P acceleration of density matrix, P d 2P dt 2 μ d 2R M 2 dt 1/ 2 E 1/ 2 P P μ P R E R Force on P P Classical dynamics in {{R,V,M},{P,W,m}} phase space Solutions obtained using velocity Verlet integrator m effects an adjustment of time-scales: Direction of Increasing Frequency of m : P changes slower with time: characteristic frequency adjusted But Careful - too large m: non-physical Appropriate m: approximate BO dynamics Consequence Bounds for m: From a Hamiltonian formalism m: also related to deviations from the BO surface “Physical” interpretation of m : Commutator of the electronic Hamiltonian and density matrix: bounded by magnitude of m F, P F 1 1/4 1/4 2 Tr μ Wμ P, W F Magnitude of m : represents deviation from BO surface m acts as an “adiabatic control parameter” Bounds on the magnitude of m The Lagrangian 1 1 T L Tr V MV Tr μ1/4Wμ1/4 2 2 E(R, P) Tr ΛP 2 2 P The Conjugate Hamiltonian (Legendre Transform) 1 1 T H Tr V MV Tr μ1/4Wμ1/4 2 2 H H real H fict : E(R, P) Tr ΛP 2 2 P dH real dH fict 1/2 dW 1/2 Tr Wμ μ dt dt dt By controlling m: control deviations from BO surface and adiabaticity Nuclear Forces: What Really makes it work Hellman-Feynman contributions E(R i , Pi ) R P dh ~ 1 dG ~ E xc VNN Tr P P 2 dR R R dR ~ Tr P, F Pulay’s moving basis terms ~ dS ~ Tr FP P dR ~ dU 1 ~ T dU T Q U - PU dR dR Contributions due to [F,P] 0. Part of non-Hellman-Feynman S=UTU, Cholesky or Löwdin Some Advantages of ADMP ADMP: – Currently 3-4 times faster than BO dynamics – Improvements will allow ADMP ~ 10 times faster – Computational scaling O(N) – Hybrid functionals (more accurate) : routine – Smaller m : Greater adiabatic control – QM/MM: localized bases: natural Comparison with BO dynamics Born-Oppenheimer dynamics: • Converged electronic ADMP: • Electronic state states. • Approx. 8-12 SCF • cycles / nuclear config. dE/dR not same in both methods • • propagated classically : no convergence reqd. 1 SCF cycle : for Fock matrix -> dE/dP Current: 3-4 times faster. 10 times Reference… H. B. Schlegel, S. S. Iyengar, X. Li, J. M. Millam, G. A. Voth, G. E. Scuseria, M. J. Frisch, JCP, In Press. Comparison with Car-Parrinello : Slide 0 Atom-centered Density Matrix Propagation (ADMP) approach using Gaussian basis sets • • Atom-centered Gaussian basis functions – Fewer basis functions for molecular systems Electronic Density Matrix propagated – Asymptotic linear-scaling with system size Car-Parrinello (CP) method • • Orbitals expanded in plane waves Occupied orbital coefficients propagated – O(N3) computational scaling References… CP: R. Car, M. Parrinello, Phys. Rev. Lett. 55 (22), 2471 (1985). ADMP: H. B. Schlegel, J. Millam, S. S. Iyengar, G. A. Voth, A. D. Daniels, G. E. Scuseria, M. J. Frisch, JCP, 114, 9758 (2001). S. S. Iyengar, H. B. Schlegel, J. Millam, G. A. Voth, G. E. Scuseria, M. J. Frisch, JCP, 115,10291 (2001). Comparison with Car-Parrinello : Slide 1 Plane-wave CP: ADMP: • Computational scaling O(N3) • Pure functionals (e.g. BLYP) – Computational scaling O(N) Hybrid (B3LYP): expensive Adiabatic control limited : larger m : D2O for H2O Properties depend on m § accurate) : routine – Smaller m : Greater adiabatic control: can use H2O – Properties independent of m # • • – Hybrid functionals (more References… § Scandolo and Tangney, JCP. 116, 14 (2002). # Schlegel, Iyengar, Li, Millam, Voth, Scuseria, Frisch, JCP, 117, 8694 (2002). Comparison with Car-Parrinello : Slide 2 Plane-wave CP: • Larger no. of basis fns. • QM/MM: Plane-waves • enter MM region Pseudopotentials required for core ADMP: • Fewer basis fns. • QM/MM: localized • bases: natural Pseudopotentials not required for core – Important for metals e.g., redox species and enzyme active sites Propagation of P: a time-reversible propagation scheme Velocity Verlet propagation of P 1/ 2 t 2 1/ 2 E(R i , Pi ) Pi 1 Pi Wi t μ i Pi Pi i i μ 2 Pi R Propagation of W 1/ 2 t 1/ 2 E(R i , Pi ) Wi 1/2 Wi - μ i Pi Pi i i μ 2 R Pi 1/ 2 t 1/ 2 E(R i 1 , Pi 1 ) Wi 1 Wi 1/2 - μ i 1Pi 1 Pi 1 i 1 i 1 μ 2 Pi 1 R Classical dynamics in {{R,V},{P,W}} phase space i and i+1 obtained iteratively: – Conditions: Pi+1 2 = Pi+1 and WiPi + PiWi = Wi Idempotency: To obtain Pi+1 Given Pi2 = Pi, need to find indempotent Pi+1 Guess: t 2 1/ 2 E(R i , Pi ) 1/ 2 * Pi 1 Pi Wi t - 2 μ Pi μ R Or guess: Pi1* Pi 2Wi t - Wi-1/2 t Iterate Pi+1 to satisfy Pi+12 = Pi+1 Pi 1 Pi 1 μ 1/ 2 Pi TPi Qi TQ i μ 1/ 2 ~ * T μ1/ 2 Pi 1 Pi 1 μ1/ 2 * Rational for choice PiTPi + QiTQi above: i Pi Pi i i Pi i Pi Qi i Qi Idempotency: To obtain Wi+1 Given WiPi + PiWi = Wi, find appropriate Wi+1 Guess: * i 1 W t 1/ 2 E(R i 1 , Pi 1 ) 1/ 2 Wi 1/2 - μ μ 2 Pi 1 R Wi 1 W * i 1 μ 1/ 2 ~ ~ Pi 1TPi 1 Qi 1TQi 1 μ 1/ 2 ~ ~ * T μ1/ 2 Wi 1 Wi 1 μ1/ 2 Iterate Wi+1 to satisfy Wi+1Pi+1 + Pi+1Wi+1 = Wi+1 Density Matrix Forces: McWeeny Purified DM (3P2-2P3) in energy expression to obtain Use E(R i , Pi ) 3FP 3PF 2FP2 2PFP 2P 2 F P R Nuclear Forces: What Really makes it work Hellman-Feynman contributions E(R i , Pi ) R P dh ~ 1 dG ~ E xc VNN Tr P P 2 dR R R dR ~ Tr P, F Pulay’s moving basis terms ~ dS ~ Tr FP P dR ~ dU 1 ~ T dU T Q U - PU dR dR Contributions due to [F,P] 0. Part of non-Hellman-Feynman S=UTU, Cholesky or Löwdin Idempotency (N-Representibility of DM): Given Pi2 = Pi, need i to find idempotent Pi+1 Solve iteratively: Pi+12 = Pi+1 Given Pi, Pi+1, Wi, Wi+1/2, need i+1 to find Wi+1 Solve iteratively: Wi+1 Pi+1 + Pi+1 Wi+1 = Wi+1 How it all works … Initial config.: R(0). Converged SCF: P(0) Initial velocities V(0) and W(0) : flexible P(t), W(t) : from analytical gradients and idempotency Similarly for R(t) And the loop continues… Results For Comparison with Born-Oppenheimer dynamics • Formaldehyde photo-dissociation • Glyoxal photo-dissociation New Results for Protonated Water clusters Protonated water wire Ion transport through gramicidin ion channels Protonated Water Clusters Important systems for: • Ion transport in biological and condensed systems • Enzyme kinetics • Acidic water clusters: Atmospheric interest • Electrochemistry Experimental work: • Mass Spec.: Castleman • IR: M. A. Johnson, M. Okumura • Sum Frequency Generation (SFG) : Y. R. Shen, M. J. Schultz and coworkers Variety of medium-sized protonated clusters using ADMP Protonated Water Clusters: Hopping via the Grotthuss mechanism True for 20, 30, 40, 50 and larger clusters… (H2O)20H3O+: Magic number cluster Hydronium goes to surface: 150K, 200K and 300K: B3LYP/6-31+G** and BPBE/6-31+G** Castleman’s experimental results: • 10 “dangling” hydrogens in cluster • – Found by absorption of trimethylamine (TMA) 10 “dangling” hydrogens: consistent with our ADMP simulations But: hydronium on the surface Larger Clusters and water/vacuum interfaces: Similar results Predicting New Chemistry: Theoretically A Quanlitative explanation to the remarkable Sum Frequency Generation (SFG) of Y. R. Shen, M. J. Schultz and coworkers Protonated Water Cluster: Conceptual Reasons for “hopping” to surface Hydrophobic and hydrophillic regions: Directional hydrophobicity (it is amphiphilic) H3O+ has reduced density around Reduction of entropy of surrounding waters Is Hydronium hydrophobic ? H2O coordination 4 H3O+ coordination =3 Spectroscopy: A recent quandry Water Clusters: Important in Atmospheric Chemistry Bottom-right spectrum From ADMP agrees well with expt: dynamical effects in IR spectroscopy Explains the experiments of M. A. Johnson Experimental results seem to suggest this as well Y. R. Shen: Sum Frequency Generation (SFG) • IR for water/vapor interface shows dangling O-H bonds • intensity substantially diminishes as acid conc. is increased • Consistent with our results – Hydronium on surface: lone pair outwards, instead of dangling O-H • acid concentration is higher on the surface Schultz and coworkers: acidic moieties alter the structure of water/vapor interfaces References… P. B. Miranda and Y. R. Shen, J. Phys. Chem. B, 103, 3292-3307 (1999). M. J. Schultz, C. Schnitzer, D. Simonelli and S. Baldelli, Int. Rev. Phys. Chem. 19, 123-153 (2000) Protonated Water Cluster: Conceptual Reasons for “hopping” to surface Hydrophobic and hydrophillic regions: Directional hydrophobicity H3O+ has reduced density around Reduction of entropy of surrounding waters Is Hydronium hydrophobic ? H2O coordination 4 H3O+ coordination =3 Protonated Water Clusters: progress of the proton Most protonated water closer to the surface as simulation progresses 3 ang Protonated Water Cluster: Radial Distribution Functions O*-O Radial Distribution function peaks: • • BLYP : ~2.45 Angstrom and ~2.55 Angstrom B3LYP : ~2.45 Zundel [H5O2+]: ~2.45 Eigen [H9O4+]: ~2.55 BLYP : Zundel and Eigen B3LYP: Zundel BLYP : proton more delocalized Protonated Water Wire Proton hopping across “water wire” • Model for proton transfer in: – ion channels – Enzymes – liquids DFT - B3LYP / 6-31+G** / 300K / ~1 ps Basis set / functional: good water-dimer properties Protonated Water Wire Protonated Oxygen peak ~ 2.4 Angstrom Non-protonated Oxygen peaks : spread (about 2.8 Ang.) Results consistent with Brewer, Schmidt and Voth using EVB model Water wire to Ion Channels: QM/MM ADMP Proton transport through ion-channel QM/MM AIMD approach to QM/MM treatment of bio-systems EE MM full E QM I E MM I I Unified treatment of the full system within ADMP ONIOM: Energy partitioning EE MM II E QM QM QM E QM F H H I I I I, self MM E MM E I I, self QM I E MM Z j j r R j MM I MM QM Z Z i j j i Ri R j MM QM Z Z i j j i R R i j Link atom coordinates are expressed in terms of their neighbors: Link atoms factor out Preliminary results: Side-chain contributions to hop: B3LYP and BLYP: qualitatively different results Protonated Water Cluster v/s Protonated Water Wire Cluster: Proton goes to surface Wire: Proton tends to center Why? Cluster: • H3O+ coordination number 3 • Lone pair has reduced water density around Wire: • • • 2 H-bonds at center: 1 H-bond at end H3O+ lone pair has reduced density: center and edge Reduced density not a factor: Number of H-bonds is HCHO photodissociation Photolysis at 29500 cm-1 : To S1 state • Returns to ground state vibrationally hot • Product: rotationally cold, vibrationally excited H2 • And CO broad rotational distr: <J> = 42. Very little vib. Excitation H2CO H2 + CO: BO and ADMP at HF/3-21G, HF/6-31G** Glyoxal 3-body Synchronous photofragmentation What about BSSE? Due to: • • difference in instantaneous incompleteness in basis set. Atom centered nature of basis set (not present in planewaves). Worst when neighbouring atoms leave completely (ie, total dissociation). Present case: proton hopping, no complete dissociation (replaced by new proton). Expected to be less. Dominant sources of errors: • • Off the BO surface DFT functional What about BSSE? Difference in completeness of basis set. Worst when neighbouring atoms leave completely (ie, total dissociation). Dynamics without total dissociation: • Effect expected to be less. Dominant sources of errors: • DFT functional Chloride-Water Cluster Conservation Properties : 1 1/4 1/4 2 Fictitious KE = Tr μ Wμ 2 Change in Fict. KE ~ 0.0002% of total Energy Chloride-Water Clusters: Red-shifts Bend: ~ 1600 cm-1, Stretch ~3400 & ~3600 cm-1 Exptal. O-H Red Shift for Cl- (H2O)1 : – 3130 cm -1 Ar matrix : M. A. Johnson, Yale University – 3285 cm -1 CCl4 matrix : M. Okumura, CalTech Critical to use hydrogens in these simulations DFT – B3LYP / 6-31G* Chloride-Water Cluster: Cl- (H2O)25 ADMP dynamics oscillates about the BO result. Protonated Water Cluster: IR Spectrum Bending ~ 1600-1700 cm-1. Stretch: broad: 3000 – 3700 cm-1. Libration modes at less than 800 cm-1 Broad Stretching band: due to proton affecting the H-bond network Conclusions ADMP is powerful new approach to ab initio molecular dynamics • Linear scaling with system size • Hybrid (more accurate) density functionals • Smaller values for fictitious mass allow – treatment of systems with hydrogens is easy (no deuteriums required) – greater adiabatic control (closer to BO surface) Examples method bear out the accuracy of the