Scientific, Medical, and Social Worlds › Molecular, Climate, and Materials Dynamics

Classical molecular dynamics provides a concrete mechanistic state transition. In the idealized closed, classical system considered here, the state consists of atom positions and momenta. A differentiable potential $U$ determines forces, and a numerical integrator advances that state. Initial conditions and $U$ determine an ideal trajectory when the equations have a unique solution; numerical trajectories approximate it. Thermostats, solvent approximations, external forces, and reactions require additional modeling choices. A molecular simulation is not automatically an autonomous or exact representation of an experiment.

For $N$ atoms in three dimensions, positions and momenta are each $N\times3$ arrays. For atom $i$, mass $m_i$ times its acceleration equals the negative spatial gradient of potential energy:

$$m_i \frac{d^2 \mathbf{r}_i}{dt^2} = -\nabla_{\mathbf{r}_i} U(\mathbf{r}_1, \ldots, \mathbf{r}_N)$$

The left-hand side is mass times acceleration; the right-hand side is force. For example, a harmonic potential $U(r)=kr^2/2$ gives a restoring force $-kr$. This equation defines dynamics under the chosen potential, not the accuracy of that potential for every molecular configuration.

where:

• $m_i$: the mass of atom $i$
• $\mathbf{r}_i$: the position vector of atom $i$
• $U(\mathbf{r}_1, \ldots, \mathbf{r}_N)$: the potential energy surface that maps a configuration of all $N$ atoms to a scalar energy
• $-\nabla_{\mathbf{r}_i} U$: the force on atom $i$, given by the negative gradient of the potential energy with respect to its position.

Determinism and practical predictability are different. In a chaotic dynamical regime, initially nearby trajectories can separate rapidly. The rate depends on the system and on the measured separation; there is no single picosecond horizon for every liquid. Uncertain initial conditions, an approximate potential, and finite-step integration each affect prediction. Statistical properties may remain useful even when a particular microscopic trajectory is no longer accurately tracked.

Two features of this setting matter for the world-model framing.

First, time-scale separation can be expensive. If an illustrative calculation uses a one-femtosecond step, a microsecond requires $10^9$ steps and a millisecond requires $10^{12}$. These are nine and twelve orders of magnitude in time, respectively. The appropriate step depends on the fastest retained motions and any constraints imposed; one femtosecond is not universally required. Long-timescale events motivate enhanced sampling, coarse-graining, and learned surrogates. More compute can extend direct simulations, but it does not remove the cost or establish that rare events were adequately sampled.

Second, the potential and the integration procedure must both fit the question. Empirical force fields approximate bonded and nonbonded interactions; electronic-structure calculations provide another reference whose accuracy depends on their own approximations. A learned potential $U_\theta$ approximates reference energies and forces, often with cheaper evaluation after training. Cost, accuracy on relevant configurations, and reference-data coverage are separate criteria. A fast potential that fails near a transition state can give misleading kinetics even if its average test error is small.

Symmetry-aware potentials connect to the representations in Part III. For an isolated system without an external directional field, energy should not depend on rigid translation or rotation; forces should rotate with the coordinates. Relabeling identical atoms should not change physical predictions. Local energy decompositions and differentiability are additional modeling choices: long-range interactions may need explicit treatment, and a gradient can still be inaccurate.

The NequIP study by Batzner and colleagues builds equivariant internal features and obtains forces as gradients of predicted energy. It reports data-efficiency improvements on specified benchmarks. That supports this architecture in those tests, not a universal ranking over every representation or deployment regime. Useful design requirements include:

• Permutation invariance. Relabeling identical atoms should not change the predicted energy.
• Translation and rotation invariance. For the isolated system considered here, energy is invariant and forces rotate with coordinates.
• Locality. A neighborhood-based approximation can reduce evaluation cost; electrostatic or other long-range contributions may require a separate mechanism.
• Smoothness. An energy-gradient force model requires differentiability. Smoothness alone does not guarantee accurate forces or stable finite-step integration.

Enforcing a valid symmetry removes the need to learn its coordinate dependence from examples. It can prevent one kind of error even in unseen orientations. Whether it improves overall accuracy or sample efficiency depends on the task, representation, optimization, and correctness of the assumed symmetry. An external field or asymmetric boundary can change the transformations a model should respect.

Structure prediction and dynamical simulation answer different questions. AlphaFold2 predicts protein coordinates and confidence from sequence-related inputs, including evolutionary information. Its structure prediction does not supply a calibrated time-transition law for folding, conformer populations, or ligand-binding kinetics. A predicted structure can be an input to further calculations; those calculations require their own dynamics and validation. This does not establish that structures contain no information about stability or mutation effects. It establishes that a coordinate prediction alone is not a validated kinetic or intervention forecast.

Materials discovery also separates candidate generation from validation. GNoME uses graph-network predictions and reference calculations to search candidate crystal structures. Predicted thermodynamic stability against selected competing phases is a screening criterion, not proof of a synthesis route. Sun and colleagues study observed metastable inorganic phases; ground-state stability is not a necessary condition for every realizable material. Kinetic accessibility, precursor chemistry, temperature, and processing conditions matter. A static stability score does not answer all of those questions.

Weather modeling uses spatial fields such as temperature, winds, humidity, and pressure. A discretized atmospheric state might stack $C$ variable-level channels on a latitude-longitude grid, giving a $C\times H\times W$ array, although spectral and unstructured grids use other representations. Numerical dynamics and learned forecast operators both advance this represented state; observations need not measure it completely.

The state vector advances forward at each time step through a forecast operator:

$$\mathbf{x}_{t+\Delta t} = \mathcal{M}(\mathbf{x}_t)$$

where:

• $\mathbf{x}_t$: the atmospheric state vector at time $t$, containing the discretized fields on the grid
• $\mathcal{M}$: the forecast operator that advances the state forward by one time step $\Delta t$
• $\Delta t$: the forecast time step

Weather forecasting first estimates the initial state from observations, often by data assimilation, and then advances a forecast model. This reuses the filtering perspective from Part II. Chaotic error growth limits detailed trajectory prediction, but skill depends on variable, spatial scale, lead time, initial uncertainty, and metric. The familiar roughly two-week discussion is not an exact ceiling after which every weather quantity has zero information. Lorenz's multiscale analysis explicitly distinguishes predictability at different scales.

A learned model can estimate the forecast operator directly from reanalysis examples. GraphCast's 2023 report demonstrates ten-day global forecasting and reports performance against specified deterministic numerical baselines. The computation and verification targets are part of the result: a benchmark win does not establish every variable, lead time, uncertainty product, or climate application. Fast evaluation can make more ensemble or sensitivity runs affordable, provided the ensemble design itself represents relevant uncertainty.

Weather and climate queries differ in both horizon and estimand. A weather query asks about a particular evolving state; a climate query may ask about the distribution of weather under specified forcing and boundary conditions. Long-term forced change requires a response to changing drivers and relevant coupled components. The IPCC assessment of initialized decadal prediction describes contributions from initial conditions and internal variability, so climate is not simply a boundary-value problem with an irrelevant initial state.

A model fitted to historical weather transitions is constrained by data from the historical forcing regime; fitting alone is not validation, which needs a separate stated test. If an action-like forcing is absent or nearly constant in its input data, those data do not establish how its response should change when the forcing changes. That is an identification and validation concern, not a theorem that a numerical derivative must be zero. A learned emulator trained across forcing scenarios may have different evidence from a weather-only predictor.

The 2024 NeuralGCM paper combines a differentiable atmospheric dynamical core with learned physical tendencies. A schematic continuous-time description is

$$\frac{d\mathbf x}{dt}=f_{\mathrm{core}}(\mathbf x,F_t)+f_\theta(\mathbf x,F_t,z_t).$$

Here $\mathbf x$ is the modeled atmospheric state, $F_t$ denotes prescribed forcing, and $z_t$ is optional stochastic input. The core and learned module provide rates of change; a time integrator advances the state. This schematic is not a next-state sum of a core prediction and an MLP output.

The reported multidecade simulations prescribe sea-surface temperature and sea ice; they are not unrestricted coupled future-climate projections. The paper also reports failure to extrapolate to substantially different future climates. A resolved dynamical core makes the division of processes inspectable, but learned source terms and numerical integration still require budget, stability, and regime testing. Neither the architecture nor a stable run proves exact conservation or causal validity.

Physical constraints are useful when they apply to the actual system and boundary conditions. A symmetry can enforce coordinate consistency without establishing the potential's accuracy. Conservation of an isolated continuous-time model does not establish accuracy of its trajectories or exact conservation by a numerical integrator. Driven and open systems need the appropriate sources, sinks, and fluxes rather than a blanket constant-energy rule. Representation quality, numerical quality, and empirical predictive fidelity remain separate questions.