Glossary

Full forms for the acronyms used across the site. Each one also appears as a hover tooltip on its first use in the text. Terms link to the research thread where they are used most.

The curated terms below double as the inline hover tooltips across the site. Further down, “More terms, from the thesis” carries the full set ingested from my PhD thesis glossary, for reference and cross-linking.

PES potential energy surface #: The energy of a system as a function of its atomic coordinates. Minima are stable structures; first-order saddle points are transition states. Nearly every method on this site is a way of searching a PES cheaply.
NEB nudged elastic band #: A double-ended saddle search: a chain of replicas (“images”) between a known reactant and product is relaxed under spring coupling to trace the minimum energy path. The climbing-image variant pushes the highest image onto the saddle. related research →
CI-NEB climbing-image nudged elastic band #: The variant of NEB in which the highest-energy image climbs against the band to converge directly onto the first-order saddle point. related research →
MMF minimum mode following #: A single-ended saddle search that needs only a starting geometry. It walks uphill along the softest Hessian eigenmode (the minimum-curvature direction), which the dimer method estimates from two force evaluations per step. related research →
MEP minimum energy path #: The lowest-energy route connecting two minima on the PES, passing through the saddle point that sets the reaction barrier. related research →
GP Gaussian process #: A non-parametric Bayesian model that places a distribution over functions. Trained on computed energies and forces, it acts as a cheap surrogate PES and reports its own predictive uncertainty. related research →
GPR Gaussian process regression #: Regression with a Gaussian process; here, fitting energies and gradients to build a surrogate energy surface that steers the next expensive calculation to where uncertainty is highest. related research →
OT-GP Optimal Transport Gaussian Process #: A GP framework that uses the Earth Mover’s Distance as its similarity metric and farthest-point sampling for data selection, keeping the surrogate affordable and stable on long runs. related research →
EMD Earth Mover's Distance #: An optimal-transport distance between distributions; used as a permutation-invariant measure of structural similarity between molecular configurations. related research →
FPS farthest-point sampling #: A greedy subset-selection scheme that repeatedly picks the point farthest from those already chosen, used to prune redundant training data for the GP. related research →
MAP maximum a posteriori #: Point estimate at the peak of the posterior distribution; used to regularize the GP hyperparameters. related research →
RFF random Fourier features #: A randomized approximation of a kernel by a finite feature map, trading a small accuracy loss for near-linear scaling of GP inference. related research →
DFT density functional theory #: The workhorse electronic-structure method that supplies the reference energies and forces. Each evaluation is expensive, which is why surrogate acceleration matters.
MLIP machine-learned interatomic potential #: A model (neural network or kernel) trained on electronic-structure data to predict energies and forces at a fraction of the cost, transferable across a class of systems. related research →
MM molecular mechanics #: Classical force-field modelling of a system from bonded and non-bonded parameters. Cheap enough for large systems and long timescales, but blind to bond breaking. related research →
ONIOM our own n-layered integrated molecular orbital and molecular mechanics #: A subtractive multi-layer embedding, E = E_MM(full) + E_high(region) - E_MM(region), that treats a small reactive region at high accuracy while a cheaper method handles the surroundings. Used here to embed an MLIP region inside an MM force field. related research →
MD molecular dynamics #: Time integration of Newton’s equations for a system of atoms; the simulation engine (e.g. GROMACS) into which MLIP/MM embedding is wired.
KMC kinetic Monte Carlo #: A stochastic method that advances a system between states by their rate constants, reaching timescales far beyond molecular dynamics.
AKMC adaptive kinetic Monte Carlo #: KMC in which the table of escape mechanisms is discovered on the fly by saddle searches rather than enumerated in advance. The long-horizon target of the surrogate-accelerated saddle work. related research →
HTST harmonic transition state theory #: The harmonic approximation to transition state theory that turns saddle and minimum geometries (plus their Hessians) into a rate constant.
RMSD root-mean-square deviation #: The root-mean-square atomic displacement between two structures after optimal alignment; used as a reaction coordinate for the two-dimensional path projections. related research →
MLIP/MM machine-learned potential embedded in molecular mechanics #: An embedding in which an MLIP drives a reactive subsystem while a classical MM force field handles explicit solvent and bulk, bringing MLIP accuracy to system sizes pure MLIPs cannot reach. related research →

More terms, from the thesis

Ingested from my PhD thesis glossary. These are page-only (no inline tooltips), and each is linkable by anchor, e.g. /glossary/#glossary-emd.

Armijo line search #: Method for step size selection in optimization to ensure sufficient decrease
CG conjugate gradient #
Cholesky decomposition #: Matrix decomposition into lower and upper triangular matrices, used for solving systems
credible interval #: Bayesian equivalent of a confidence interval; range of parameter values with given probability
DAG directed acyclic graph #
data pruning #: Technique to reduce dataset size by removing less informative or distant data points
deflated standard error #: Standard error estimate that is artificially small, typically due to ignoring data correlations or non-independence. Leads to overconfident statistical conclusions.
eigenvalue #: Scalar associated with an eigenvector; quantifies curvature along that direction
eigenvector #: Vector indicating a principal direction in a matrix, often used in optimization
extrema #: Minima, maxima, saddle points
FEM finite element method #
fixed effect #: Model term estimating the effect of predictors assumed to be constant across all groups
Gaussian Process #: A collection of random variables, any finite number of which have a joint multivariate normal (Gaussian) distribution; defines a distribution over functions.
Gaussian Process Regression #: Non-parametric Bayesian approach for regression and function approximation
Generalized Linear Mixed Model #: Statistical model incorporating both fixed and random effects
GPDimer Gaussian process regression dimer #
GPU graphics processing unit #
Gram-Schmidt process #: Procedure for orthonormalizing a set of vectors
GUI graphical user interface #
Hessian #: Second derivative matrix of a function, used to analyze curvature and stationary points
HF Hartree-Fock #
homoscedasticity #: Statistical property where the variance of residuals remains constant across all observations
HPC high performance computing #
hyperparameters #: Parameters governing model behavior but not directly optimized during training
IDPP image dependent pair potential #
IRA iterative rotations and assignments #
kernel function #: Function defining similarity between data points in GPR; determines covariance structure
KS Kohn-Sham #
L-BFGS limited-memory Broyden-Fletcher-Goldfarb-Shanno #
likelihood #: Probability of observed data given specific model parameters
marginal likelihood #: Evidence; probability of observed data under a model, integrating over parameters
MLL marginal log-likelihood #
MMF-NEB minimum mode following nudged elastic band #
multiple-input multiple output multiple-input multiple output #
MVN multivariate normal #
non-independence #: Condition where observations are correlated or related, violating the assumption of independent samples
OTGPD optimal transport Gaussian process dimer #
posterior distribution #: Probability distribution representing updated beliefs after observing data
Potential Energy Surface #: Multidimensional surface representing energy as a function of geometry
prior distribution #: Initial probability distribution before observing data
PSD positive semi-definite #
pseudoreplication #: Error in statistical inference caused by treating non-independent observations as independent
Radial Basis Function #: Common kernel function in GPR, based on distance between points
random intercept #: Model term allowing each group (e.g., a chemical system) to have its own baseline value
RKHS reproducing-kernel Hilbert space #
robust #: Describes an estimator or method that maintains performance even when assumptions are violated
RPC remote proceedure call #
saddlepoint #: Point of zero force, with a single negative eigenvalue of the Hessian
SCF self consistent field #
SCG scaled conjugate gradient #
SE squared exponential #
SPD symmetric positive definite #
standard error #: Estimate of the variability of a sample statistic, often used to quantify uncertainty in parameter estimates
TST transition state theory #
variance #: Measure of spread in a dataset or uncertainty in a model prediction
WBO Wiberg bond order #