ab initio simulation

**Etch Plasma–Surface Ab Initio Molecular Dynamics (AIMD) Modeling follows atomic trajectories while recomputing electronic-structure forces from first principles at every time step, allowing bond formation/breaking, polarization, charge redistribution, collision cascades, product formation, and short-time surface restructuring without a pre-fitted classical reactive potential.** Its defensible output is a convergence-qualified ensemble of mechanisms, forces, prompt outcome statistics, and reference configurations—not a single expensive trajectory promoted to an etch yield. This upgraded page owns the short-time dynamical bridge between static DFT and larger reactive/classical MD. Static DFT owns stationary states, thermochemistry, and saddle-point barriers; AIMD tests finite-temperature motion and prompt reactions on the chosen electronic surface; nonadiabatic/electron dynamics methods own electronic transitions when the Born–Oppenheimer assumption fails; classical or machine-learned MD owns larger impact ensembles; kMC owns rare-event waiting time; feature models own particle transport and profile evolution. | AIMD layer | Required definition and the failure it prevents | |---|---| | physical question | Material/surface state, incident species, kinetic energy/angle, temperature, charge/spin/electronic assumptions, dose and exported observable; prevents an illustrative trajectory from answering a statistical process question. | | dynamical formulation | Born–Oppenheimer, Car–Parrinello, Ehrenfest/nonadiabatic variant; nuclear/electronic equations, ensembles and conserved quantity; prevents incompatible trajectories from sharing one “AIMD” label. | | electronic method | Code/version, XC/dispersion, spin, pseudopotential/basis, cutoff/k mesh, occupation/smearing, charge and SCF/root-following settings; prevents force errors from masquerading as chemistry. | | atomic specimen | Facet/amorphous replicas, coverage, native oxide/polymer, defects/damage, lateral cell, slab/vacuum, fixed/thermal layers and preparation; prevents periodic/boundary artifacts from determining impact outcome. | | trajectory protocol | Incident sampling, launch/reference, timestep/adaptation, SCF tolerance, integrator, thermostat, run length, escape/stopping rules and checkpoints; prevents drift, premature classification and artificial heat removal. | | outcome analysis | Persistent adsorption/reflection/reaction/product/removal/implantation/damage definitions with atom, charge and energy ledgers; prevents transient motion from becoming a yield. | | statistical design | Independent thermal/surface/site/orientation replicas, weights, censored outcomes, confidence and convergence; prevents correlated femtoseconds from becoming independent evidence. | | scale-up contract | Raw configurations/forces, conditional outcomes, validity range, uncertainty and provenance for DFT/ML-MD/kMC/feature consumers; prevents uncontrolled extrapolation and double counting. | **Choose the dynamical approximation explicitly.** In Born–Oppenheimer molecular dynamics (BOMD), nuclei evolve classically on an electronic ground-state potential energy surface recomputed at each configuration: $$ M_I\ddot{\mathbf R}_I=-\nabla_{\mathbf R_I}E_{BO}(\{\mathbf R\}). $$ The electronic problem is solved self-consistently at every nuclear step, commonly with Kohn–Sham DFT, $$ \widehat H_{KS}[n;\{\mathbf R\}]\psi_i=\epsilon_i\psi_i, \qquad n(\mathbf r)=\sum_if_i|\psi_i(\mathbf r)|^2. $$ For a complete basis the force is the Hellmann–Feynman contribution plus ion–ion terms; basis dependence can add Pulay forces. BOMD assumes electrons remain on the selected adiabatic state as nuclei move. An SCF-converged step solves the chosen approximation, not necessarily the real excited/charge-transfer dynamics of an ion impact. Car–Parrinello MD propagates auxiliary electronic degrees of freedom with a fictitious mass while constraining orbital orthonormality. It can avoid full SCF minimization each step when adiabatic separation is maintained, but the conserved extended energy differs from physical nuclear energy, and fictitious electronic motion must not exchange appreciable energy with ions. Report fictitious mass, integration timestep, electronic kinetic energy, initialization and drift. Ehrenfest, surface hopping, real-time TDDFT, constrained DFT dynamics, electronic friction and related nonadiabatic methods address different electronic-transition questions. They are not interchangeable upgrades to BOMD. Define electronic states, decoherence, hopping/force rules, charge reservoir and validation; otherwise expose missing excitation/neutralization as model-form uncertainty. **Electronic forces inherit every static-DFT approximation.** State exchange–correlation functional, dispersion, exact exchange/$U$, spin polarization, relativistic treatment, pseudopotential or all-electron method, basis/cutoff, reciprocal sampling, occupations/smearing, boundary conditions and correction schemes. Benchmark choices against the chemistry and high-energy configurations encountered—not only equilibrium bulk structure. An AIMD collision may access compressed interatomic distances, unusual coordination, radicals, fragments, transient metallicity and high electronic temperature. Pseudopotential valence partition and short-range core overlap must remain valid. Compare repulsive curves/forces to harder potentials or all-electron references over the closest approaches expected. A potential designed for equilibrium solids may fail before nuclei touch. Semilocal DFT self-interaction can over-delocalize charge and alter bond breaking/barriers. Hybrids may improve localization but greatly raise trajectory cost. DFT+$U$ introduces projector/parameter dependence; dispersion matters for weakly bound precursors/products; spin state affects radicals and open-shell surfaces. Run method sensitivity on representative trajectory snapshots and decision outcomes. SCF occupations can switch as a surface becomes metallic or products form. Specify smearing/electronic temperature and whether the reported conserved quantity is free energy or extrapolated internal energy. Excessive smearing changes forces/chemistry; insufficient smearing can destabilize SCF. Converge it against trajectories and product classification. **SCF convergence is part of the integrator.** If electronic residuals vary randomly between steps, force noise heats nuclei and destroys time reversibility. Set energy/density/eigenvalue residuals tight enough that force error is small relative to physical forces and timestep truncation. Monitor iterations, residuals, magnetization, occupation and extrapolation failures at every step. Use wavefunction/density extrapolation from prior steps to accelerate convergence, but protect against following the wrong electronic root through bond breaking or spin/charge rearrangement. Periodically restart from less biased initial guesses and compare. A trajectory that survives only because it remains trapped in one SCF basin needs explicit interpretation. For microcanonical BOMD, monitor $$ E_{tot}(t)=\sum_I\frac12M_I|\mathbf V_I|^2+E_{BO}(\{\mathbf R(t)\}). $$ Drift and high-frequency oscillation should converge with timestep and SCF tolerance. Separate integrator truncation, SCF force error, thermostat work, boundary work, external-field work and intentional electronic stopping. A flat plotted temperature can hide large unreported thermostat energy. **Build a plasma-facing surface ensemble.** Specify crystalline orientation/reconstruction or produce multiple independent amorphous structures with qualified density, composition, coordination and stress. Include process-relevant halogen/hydrogen/oxygen/carbon coverage, native oxide, polymer, vacancies, implanted atoms, roughness and damage. Equilibrate each surface at target temperature using a declared thermostat/ensemble, then draw decorrelated positions and Maxwell–Boltzmann velocities. Check energy, temperature by region, stress, coordination and composition. Consecutive frames separated by a few femtoseconds are not independent surface replicas. Use lateral periodic cells large enough that collision cascades, polarization, fragments and strain fields do not interact with images. Converge outcome-sensitive cell size; a projectile repeatedly sees its image-defined coverage/site pattern. The slab must be thick enough to isolate the active region from fixed/bottom boundaries during the analysis window. Vacuum must accommodate launch, reflection, clusters and product classification without interaction across the repeated normal direction. Asymmetric and charged slabs need dipole/electrostatic handling. Inspect planar charge/potential and density in vacuum. An escaping electron or charged fragment in periodic DFT is not automatically a physical open boundary. A practical slab may contain fixed support atoms, thermostatted heat-sink atoms and an upper Newtonian impact zone. Converge each thickness. Do not thermostat the active collision region: it suppresses cascade energy, products and activated rearrangement. Momentum reflected from fixed atoms or phonons returning from the bottom can change late outcomes. For amorphous low-$k$, oxide and polymer materials, configuration variability is often larger than numerical error. Sample distinct local motifs and impact positions. Report the distribution; one nanopore, Si–CH$_3$ group, F-rich site or strained bond cannot represent the material. **Initialize incident conditions from the upstream plasma model or a designed beam study.** Condition histories on species $s$, charge/electronic assumption, kinetic energy $E$, direction $\Omega$, impact position, molecular orientation/internal state, surface state $\chi$ and temperature $T_s$. Preserve energy–angle correlation when using sheath distributions. For projectile mass $m_p$, $$ v_p=\sqrt{\frac{2E}{m_p}}. $$ Transform the direction relative to the local macroscopic surface normal and state the angular measure. Sample lateral coordinates over the physical cell; use symmetry only if surface composition and adsorbates possess it. Sample open-shell orientation/spin deliberately. Launch where interaction with the slab is negligible under the chosen boundary/electrostatics, or define and subtract the long-range reference. Check initial force and potential energy. Too-low launch injects an arbitrary interaction; too-high launch wastes scarce AIMD steps. An incident plasma ion is not fully defined by adding/removing one electron from a periodic supercell. Near-surface neutralization, image charge, electron emission, substrate conduction and sheath current require an electron reservoir/open-system treatment beyond ordinary fixed-electron BOMD. Declare whether the trajectory models a neutralized projectile, fixed total charge, constrained charge localization, or another ensemble. Compare plausible charge/spin preparations where they affect mechanism. Track density differences and multiple charge analyses as diagnostics, but do not call a partitioned Bader/Hirshfeld number an observed charge-transfer probability. If electron exchange controls the decision, use a qualified nonadiabatic/embedding/constant-potential approach or stop. **Choose the nuclear timestep for the hardest collision.** An equilibrium timestep can fail when an energetic projectile approaches a nucleus. Test fixed small steps or a verified reversible/adaptive strategy based on maximum force, acceleration, displacement or energy error. Variable stepping changes integration properties and must not bias outcome statistics. Velocity Verlet has local error controlled by $\Delta t$, but energy stability is empirical for the coupled SCF trajectory. Converge trajectory classifications, outgoing energy and deposited energy against timestep—not only average temperature. Ensure neighbor/projector grids and SCF extrapolation update consistently after a shortened step. An energy-based adaptive bound might require $$ \max_I|\mathbf V_I|\Delta t<\delta R_{max}, $$ along with acceleration and electronic convergence tests. Record every accepted/rejected step and reconstruct physical time exactly. Never compare per-step reaction frequency when timesteps differ. Use a thermostat only to prepare temperature or represent distant heat removal. For the prompt impact window, NVE dynamics in the active region is generally easiest to audit. If Langevin, Nosé–Hoover or boundary damping remains active, report work and show impact outcome convergence to coupling strength/location. Estimate acoustic return time from slab thickness and sound speed; classify prompt outcomes before echoes or enlarge/absorb the boundary. Electronic energy transfer not represented by ground-state DFT must not be silently absorbed into a thermostat. Maintain explicit unresolved reservoirs. **AIMD time is exceptionally short and computational flux exceptionally high.** Typical trajectories span picoseconds to tens of picoseconds, while experimental arrivals, diffusion and desorption can be microseconds or longer. Observing no event within 5 ps gives a censored trajectory, not zero rate. If cell area is $A$ and $N_{imp}$ impacts are applied, fluence is $$ \Phi=\frac{N_{imp}}{A}. $$ Mapping to time as $t=\Phi/\Gamma$ exposes that sequential AIMD shots often represent enormous artificial flux. Cascades may overlap; radicals/products have no physical replenishment/removal; heat and damage accumulate; slow chemistry is skipped. Do not call sequential impacts a reactor-time simulation without a bridging method. Use reset-surface ensembles to estimate conditional prompt outcomes at fixed $\chi$. Use cumulative bombardment only for explicitly dose-dependent structural evolution, with independent replicas, equilibration/slow-event policy, inventories and finite-reservoir controls. Alternate AIMD/MD impacts with kMC or a validated reservoir model for slow intervals. Enhanced-sampling methods—metadynamics, umbrella sampling, adaptive bias, blue-moon constraints, accelerated dynamics—can reveal free-energy barriers but alter trajectory probabilities and time. State collective variables, bias, reweighting and convergence. Biased paths cannot be inserted into an unbiased impact kernel without correction. **Classify persistent physical outcomes, not snapshots.** Define analysis/escape planes, bonding or cluster rules, persistence time, direction and retained depth. Outcomes include reflection, adsorption, dissociation, reaction, product creation/desorption, physical/chemical removal, implantation, mixing and damage. Reflection records outgoing species, energy, angle, spin/charge assumption and changed surface state. Adsorption requires stable binding over the qualified observation window or an explicitly censored label. Product formation and product escape are distinct. A fragment crossing a plane and returning must not be counted twice. Physical sputter yield counts substrate atoms/formula units removed primarily by momentum transfer; chemical etch yield counts volatile target-containing reaction products. State the unit. Yield can exceed one and is not a probability. With outcome multiplicity $n_p^{(j)}$ and history weight $w_p$, $$ \widehat Y_j=\frac{\sum_pw_pn_p^{(j)}}{\sum_pw_p}. $$ Track immutable atom identities and balance every element: $$ \mathbf N_{slab,0}+\mathbf N_{incident}=\mathbf N_{retained}+\sum_j\mathbf N_{out,j}. $$ Also ledger incident kinetic/internal energy, electronic/ionic potential change, outgoing kinetic/internal energy, lattice energy, thermostat/boundary/external work and numerical residual. Charge bookkeeping follows the declared electronic ensemble; do not infer emitted current when electrons cannot leave the cell. Damage metrics may include coordination, vacancies/interstitials, bond scission, mixing, carbon depletion, densification and residual strain after a defined relaxation. High-temperature transient coordination is not stable damage. Compare to a thermal control trajectory with no projectile. **One trajectory demonstrates possibility, not probability.** Independent variables include thermal velocities, atomic surface replica, local impact site, projectile orientation, energy/angle, charge/spin initialization and electronic-method uncertainty. Plan an ensemble or use AIMD as targeted mechanistic/reference evidence for a cheaper model. For binary outcomes, report confidence intervals and zero-event upper bounds. For yields/products, report sample variance/covariance and heavy tails. Time steps within one trajectory and multiple products from one cascade are correlated; the independent unit is usually the prepared history/surface replica. Converge separate axes: electronic method/SCF, timestep, cell/slab/vacuum, thermostat/boundary, trajectory duration, initial surface ensemble, impact sites/orientations and number of histories. A large statistical ensemble with one biased functional/cell remains precisely biased. Use sequential design: pilot diverse conditions, identify mechanism/outcome uncertainty, then allocate AIMD to decision-sensitive or potential-extrapolative regions. Importance sampling needs weights if estimating physical averages. Preserve all failures and censored runs in the denominator according to a predefined rule. **AIMD is often most valuable as training and validation data.** Export structures, energies, forces, stresses, spin/charge diagnostics and event labels from equilibrium, reaction, collision-compressed, product and damaged configurations. Sampling every adjacent timestep overweights nearly identical frames; cluster/thin by descriptor or select informative frames. For a machine-learned potential trained on reference configurations $c$, a generic loss is $$ \mathcal L=\sum_c\left[w_E|E_c-E_c^{ref}|^2+w_F\sum_I\|\mathbf F_{Ic}-\mathbf F_{Ic}^{ref}\|^2+w_\sigma\|\boldsymbol\sigma_c-\boldsymbol\sigma_c^{ref}\|^2\right]. $$ Split validation by whole trajectory/configuration family, not random neighboring frames. Hold out impact energies, products, surface states and reaction families. Validate energy conservation and stable long MD, not only static RMSE. Use active learning with committee disagreement, descriptor distance or extrapolation metrics to request new AIMD frames. Calibrate the trigger against true held-out force/energy error. Stop classical/ML trajectories on dangerous extrapolation rather than accepting chemically impossible products. Delta learning may correct a cheaper electronic level toward a higher one; record baseline/correction domains and ensure force consistency. Training to approximate DFT inherits its functional, charge and nonadiabatic errors. Challenge decisive mechanisms against higher-level theory and experiment. An ML/reactive potential can run thousands of impact replicas at larger size; AIMD should audit representative raw trajectories, mechanism ordering, force regions, outcome kernels and out-of-domain cases. Disagreement is evidence to refine the dataset or validity mask, not to tune post hoc yield multipliers. **Export scale-aware closures.** Feature Monte Carlo may consume a conditional product/reflection kernel $$ K_j(s',E',\Omega',\mu\mid s,E,\Omega,\chi,m,T_s), $$ whose integral is probability or expected multiplicity. AIMD alone rarely samples this high-dimensional kernel densely, so combine it hierarchically with ML/reactive MD and beam data. Preserve energy–angle–species correlation and uncertainty. Surface kMC consumes prompt state transitions plus thermal events. Define a commitment time separating impact dynamics from slow diffusion/desorption/reaction. Map retained atoms, coverage, damage and products conservatively. Do not execute the same prompt reaction in AIMD and later again in kMC. Static DFT/NEB should replace brute-force AIMD waiting for rare thermal events. AIMD can test finite-temperature recrossing and discover paths; enhanced sampling can estimate free energy; kMC advances qualified rates. Each rate needs state, site degeneracy, prefactor, uncertainty and validity. Feature/profile conversion requires absolute incident flux and material counting volume. AIMD yields do not contain physical arrival time. For target-unit density $n_m$, planar recession from yield $Y_m$ and flux $\Gamma$ is $$ V_n=-\frac{Y_m\Gamma}{n_m}, $$ with the same atom/formula-unit convention. Mixed layers need composition/density state. Pass surface products, heat and damage to the correct consumer once. **Nonadiabatic boundaries must be visible.** BOMD assumes electrons adjust instantaneously on one potential surface. Energetic plasma impacts can cause electron–hole pairs, electronic stopping, projectile neutralization, Auger/secondary-electron emission, excited fragments and radiation chemistry. Ground-state force trajectories cannot quantify these automatically. Compare nuclear kinetic energy and material electronic scales; inspect avoided crossings, occupation changes, charge localization and experimental evidence. Use real-time TDDFT, constrained DFT, fewest-switches surface hopping, electronic friction, GW/BSE or open-system methods only within their qualified regime. Each introduces new approximations and usually smaller feasible ensembles. If electronic stopping is added empirically to nuclei, tally removed work, specify energy/velocity/domain, and ensure it is not double counted by the electronic method. If an ion is assumed neutralized at a dividing plane, document the plane and sensitivity. Do not label a fixed-electron periodic simulation “charge-transfer resolved.” Excited-state AIMD may require tracking state identity across crossings. Root flipping can create discontinuous forces. Demonstrate state-tracking/decoherence/time-step convergence and compare against known scattering or spectroscopy. When unavailable, bound the resulting model-form uncertainty in the downstream prediction. **Verification proves the implementation before chemistry.** Reproduce static DFT energies/forces for frozen frames; finite-difference selected forces; compare equivalent cross-code settings; test isolated atom/molecule spin; and reproduce equilibrium lattice, vibrational and surface properties. Run NVE timestep/SCF convergence on equilibrium and high-force collision cases. Verify expected energy-error scaling, zero net drift, stable momentum/center of mass, temperature distributions and thermostat work. Deliberately loosen SCF and increase timestep to ensure monitors detect failure. Test initialization: kinetic energy from velocity, direction/frame, launch interaction, thermal velocities, orientation, random seeds and charge/spin. Test boundary cases: periodic crossing, grazing trajectories, product escape/return, fixed-layer impulse and acoustic echo. Test analysis with synthetic trajectories of known products and atom balances. For Car–Parrinello, verify fictitious electronic kinetic energy and adiabatic separation. For BOMD, verify SCF/root continuity. For adaptive timesteps, reconstruct time and compare against a small fixed-step reference. For enhanced/nonadiabatic methods, reproduce their own analytic/benchmark limits. | AIMD qualification gate | Evidence and stop condition | |---|---| | dynamical scope | BOMD/CP/nonadiabatic formulation, electronic state/charge, material/state, incident domain, ensemble and requested decision are explicit. | | electronic forces | XC/spin/dispersion/pseudopotential/basis/k/occupation choices pass equilibrium, reactive and short-range challenge configurations. | | integration integrity | SCF/root, timestep/adaptation, force consistency and conserved-energy/reservoir ledgers converge for thermal and impact trajectories. | | finite specimen | Independent surfaces plus lateral size, slab depth, vacuum, fixed/thermal layers and echo time leave outputs stable. | | event analysis | Persistent outcome definitions, immutable atom IDs, products/removal/damage, charge convention and energy/element ledgers pass synthetic and real cases. | | statistical evidence | Surface/site/thermal/orientation replicas, censoring, confidence/covariance and convergence support the claimed probability or remain mechanism-only. | | electronic limitation | Neutralization, excitation, stopping and electron emission are resolved by a qualified method or exposed as model-form uncertainty. | | ML/MD handoff | Diverse raw reference frames, trajectory-family holdouts, stable-force tests, active-learning calibration and OOD failure behavior pass. | | multiscale validation | Static barriers, beam/plasma outcomes, products, damage and downstream kMC/feature observables agree within separated uncertainty. | **Validation follows mechanism to observable.** First validate electronic structure against molecular bonds/spins, surface structure, adsorption, reaction energies and available high-level calculations. Then compare beam-resolved reflection, energy loss, sputter/etch threshold, product identities, angular/energy distributions, implantation and damage under matched material/state/energy/angle. Plasma validation requires upstream flux/species distributions and dose history. Compare state-dependent surface composition, carbon loss, film density, volatile products, temperature response and damage—not only a final etch rate. Mixed-species plasma can hide compensating errors in incident flux and surface probability. Forward-model experimental filters: mass-spectrometer fragmentation/transmission, XPS depth/charging, infrared selection, ellipsometric density, microscopy threshold and beam energy spread. Align initial surface preparation and analysis time. Separate measurement, incident-distribution, electronic method, finite-cell, sampling, classifier and scale-mapping uncertainty. Use held-out material, surface state, energy/angle or product evidence after development. Calibrate a small interpretable discrepancy layer rather than retuning many electronic/impact parameters to one contour. Preserve raw AIMD, lower-cost potential and calibration contributions separately. **Performance and provenance decide whether results can be trusted later.** AIMD cost scales steeply with electrons, basis, exact exchange, k points and SCF iterations. Parallelize independent trajectories, impact conditions, surface replicas and electronic work appropriately. Report accepted qualified physical time/impacts per compute-hour, including failed SCF and censored trajectories. Checkpoint atomic positions/velocities, electronic state/wavefunctions subject to portability, integrator/thermostat variables, physical time, adaptive-step state, RNG and ledgers. Restart should reproduce the claimed deterministic path or ensemble distribution. Never silently restart from a different charge/spin root. Archive structures, cells, constraints, incident definitions, code/version, functional, pseudopotential/basis identifiers and hashes/licenses, k/cutoff/smearing/SCF, integrator/timestep, thermostat, seeds, raw outputs, trajectory/event analysis and convergence notebooks. Hash every identity-defining input and output; derived kernels cite those hashes. **A gated execution sequence is efficient because AIMD is expensive.** Freeze the decision and electronic/dynamical scope; challenge the DFT forces on equilibrium, reactive and repulsive configurations; prepare independent surfaces; qualify SCF/root, timestep, cell, boundary and outcome classifier on pilot trajectories; run designed impact/thermal ensembles; close atom/energy ledgers; quantify censoring and uncertainty; validate held-out beam/surface evidence; then release reference data or conditional outcomes to ML-MD, kMC and feature models with an explicit validity mask. Stop when electronic roots or spin switch uncontrolled; SCF residual heats nuclei; timestep, slab, images or thermostat change the mechanism; charged/ion claims lack an electron reservoir; products interact with periodic images; outcomes remain transient/censored; atom or energy ledgers fail; statistics rest on one surface/site; or ground-state dynamics omits a decision-critical excitation. More compute cannot rescue the wrong dynamical ensemble. **Safety applies to validation and computing.** Plasma/beam experiments can involve high voltage/RF, vacuum, toxic/corrosive/pyrophoric gases, reactive residues, UV, hot surfaces and stored energy. Use qualified operators, approved recipes, interlocks, monitoring, ventilation, compatible materials, purge verification, PPE and lockout/tagout. Protect licensed electronic-structure data/software, controlled process data and credentials; never embed secrets in job scripts or shared trajectory archives. **A credible Etch Plasma–Surface AIMD Model is a bounded electron–nuclear experiment.** It declares the adiabatic or nonadiabatic approximation; challenges electronic forces across the configurations actually visited; represents realistic surface and incident ensembles; converges SCF, roots, timestep, cell, boundary and thermostat; distinguishes persistent outcomes from censored short trajectories; closes atom and energy ledgers; quantifies statistical and model-form uncertainty; and exports auditable reference configurations or conditional mechanisms to the models that own larger ensembles, longer time and profile evolution. That is how first-principles dynamics becomes predictive plasma–surface evidence rather than one compelling movie.