Scientific Models and Simulations: Purpose, Types, and Limitations

Scientific models and simulations are the tools researchers use when the real thing is too large, too small, too slow, too fast, or too dangerous to study directly. This page covers what distinguishes a model from a simulation, how each type functions mechanically, where they appear across scientific disciplines, and where their built-in assumptions start to crack. Understanding these boundaries is as important as understanding the tools themselves — a model that is trusted beyond its valid range can produce confidently wrong answers.

Definition and scope

A scientific model is a simplified representation of a system or process, built to highlight specific features while deliberately ignoring others. That last part — the deliberate ignoring — is not a flaw. It is the mechanism. A road map is a model: it omits elevation, soil type, and ambient temperature because those features are irrelevant to the question of how to drive from Denver to Albuquerque.

Simulations extend models into time. Where a static model describes a system's structure, a simulation runs that structure forward (or backward) under specified conditions, generating sequences of states. The climate models run by the National Oceanic and Atmospheric Administration (NOAA) are simulations: they encode physical equations governing atmospheric dynamics and then advance those equations across decades of simulated time.

Scope matters here. Models exist across every scientific domain — molecular dynamics models in biochemistry, epidemiological compartment models in public health, finite element models in structural engineering, agent-based models in ecology. Each carries its own assumptions and each was built to answer a bounded set of questions. Applying any model outside that boundary is one of the most common failure modes in applied science.

How it works

Every model rests on three layers:

1. Governing equations or rules — the mathematical or logical relationships that define how components of the system behave (e.g., Newton's second law, rate equations, probability distributions).
2. Parameters — numerical values assigned to variables within those equations, often estimated from empirical data or calibrated against known outcomes.
3. Boundary conditions — the constraints that define the edges of the system: what is inside the model, what is outside, and what the model assumes about the outside.

Simulations add a fourth layer: a time-stepping mechanism. At each discrete interval, the simulation recalculates the state of every component based on the governing equations and the state at the previous step. The US Department of Energy's national laboratories run particle physics simulations where a single run can require millions of processor-hours on machines like the Frontier supercomputer at Oak Ridge National Laboratory, which delivered 1.102 exaflops of performance (Oak Ridge National Laboratory, Frontier system documentation).

The critical distinction between deterministic and stochastic models shapes how results are interpreted. A deterministic model produces one output for a given set of inputs — run it twice, get the same answer. A stochastic model incorporates randomness explicitly; run it 1,000 times and the output is a probability distribution, not a single value. Epidemiological models for infectious disease spread, like those used by the Centers for Disease Control and Prevention (CDC), are typically stochastic because individual transmission events are inherently probabilistic.

Common scenarios

Scientific models and simulations appear wherever direct experimentation is constrained. The scenarios fall into recognizable patterns:

• Scale impossibility: Cosmological simulations of galaxy formation span billions of light-years and billions of years — direct observation captures only a snapshot. NASA's IllustrisTNG project simulates cosmic structure formation across a volume roughly 300 megaparsecs on a side.
• Ethical constraint: Clinical trials cannot ethically expose human subjects to known toxins. Pharmacokinetic models simulate how drugs move through the body, allowing researchers to predict dosing thresholds before human trials begin.
• Cost compression: Crash-testing every vehicle configuration physically would be prohibitively expensive. Finite element simulations allow engineers to test hundreds of design variants before a single physical prototype is built. The National Highway Traffic Safety Administration (NHTSA) integrates computational simulation data into its vehicle safety evaluation processes.
• Counterfactual testing: Models allow scientists to ask "what if" questions impossible to run in reality — what if atmospheric CO₂ had remained at 280 parts per million? What if a specific gene were knocked out? The counterfactual world cannot be observed; it can only be modeled.

For a broader orientation to how evidence and method interact across scientific disciplines, the Science Reference Index provides structured access to related foundational topics.

Decision boundaries

No model is universally valid. Every model has a domain of applicability, and results degrade — sometimes catastrophically — when that domain is exceeded.

The most common failure modes:

• Overfitting: A model tuned too precisely to historical data loses predictive power on new data. It has memorized noise instead of learning signal.
• Parameter uncertainty: When parameters are estimated from sparse or biased data, small errors compound through iterative calculations. This is why ensemble modeling — running many simulations with slightly varied parameter sets — has become standard practice in climate science, as endorsed by the Intergovernmental Panel on Climate Change (IPCC) in its Sixth Assessment Report (IPCC AR6, Working Group I, Chapter 4).
• Structural assumptions: Every model assumes something about the shape of the system — that relationships are linear, that populations are well-mixed, that materials are homogeneous. When reality violates those assumptions, the model's outputs diverge from observation.
• Validation gaps: A model validated against data from one context may not transfer to another. A flood model calibrated on temperate watersheds may perform poorly in arid environments with flash-flood dynamics.

The distinction between verification and validation is precise and important. Verification asks whether the model was built correctly — does the code faithfully implement the intended equations? Validation asks whether the correct model was built — do the equations actually represent the real system? (NIST defines both in NIST SP 1500-207, Simulation Model Validation.) Conflating the two is how well-coded, badly-conceived models get trusted in high-stakes decisions.