ArchiveCSE 2022 was held on December 17–18, 2022. It is a past event: submissions and registration are closed. This website is an independent educational and historical resource.

Modelling • Numerical Methods • Simulation

Computational Science: Methods, Modelling and Engineering

How scientists and engineers turn physical laws into equations, equations into algorithms, and algorithms into results they can trust — and where that process can go wrong.

Last reviewed: · Editorial team, cse2022.org

01What Is Computational Science?

Computational science is the use of mathematical models and computer simulation to understand and predict the behaviour of natural and engineered systems.

It is often called the third pillar of science, alongside theory and experiment. Theory provides the governing equations; experiment provides observations; computation connects the two by solving equations that are too complex to solve by hand and by exploring conditions that experiments cannot easily reach.

Computational science differs from general software development and information technology in its goals. A business application succeeds if it meets user requirements. A scientific code succeeds only if its output is a faithful approximation of reality, with a known and acceptable error. The software engineering matters, but the decisive questions are mathematical and scientific.

How computational science differs from neighbouring fields
FieldMain questionTypical output
Computational scienceDoes the model predict reality accurately?Validated simulation results with error estimates
Computer scienceWhat can be computed, and how efficiently?Algorithms, proofs, systems
Software developmentDoes the software meet its requirements?Working applications
Information technologyDo systems serve the organization reliably?Operated infrastructure and services

02Relationship Between Mathematics and Computing

Most physical laws are written as differential equations, which describe how quantities change. Computers, however, can only store finite lists of numbers and perform arithmetic on them. The core task of computational science is to bridge that gap: to replace continuous mathematics with discrete approximations that a computer can handle, while proving or measuring that the approximation converges to the true answer as resolution increases.

Floating-point arithmetic adds a second layer. Computers represent real numbers with finite precision (about 16 significant decimal digits in standard double precision), so every operation can introduce a tiny rounding error. Usually these errors are harmless; in poorly conditioned problems they can grow until they dominate the result.

03Scientific Modelling

A model is a deliberate simplification. To model heat flow in a metal rod, we might assume the rod is thin enough that temperature varies only along its length, that the material is uniform, and that the ends are held at fixed temperatures. These assumptions lead to the one-dimensional heat equation:

∂u/∂t = α ∂²u/∂x²

Here u is temperature, t time, x position and α the thermal diffusivity. Every model has a domain of validity. The heat equation above is excellent for slow conduction in solids, and wrong for situations where radiation or phase changes matter.

Good modellers document their assumptions explicitly, so that others can judge whether the model applies to their own problem.

04Numerical Methods

Numerical methods are algorithms for obtaining approximate solutions to mathematical problems. Major families include:

  • Finite difference methods, which replace derivatives with differences between values at grid points.
  • Finite element methods, which divide a domain into small elements and approximate the solution within each one; widely used in structural engineering.
  • Finite volume methods, which conserve quantities such as mass and energy across cell boundaries; common in fluid dynamics.
  • Iterative solvers such as conjugate gradient, which solve large systems of linear equations without forming a full matrix inverse.
  • Monte Carlo methods, which estimate quantities by random sampling and are valuable in high-dimensional problems.

Applied to the heat equation, an explicit finite-difference scheme updates each interior grid point from its neighbours:

for each time step n:
    for i = 1 .. N-2:
        u_new[i] = u[i] + r * (u[i+1] - 2*u[i] + u[i-1])
    u = u_new          # r = alpha * dt / dx^2

This scheme is only stable when r ≤ 1/2. If the time step is too large relative to the grid spacing, errors grow exponentially and the solution becomes meaningless — a concrete example of why numerical analysis matters.

05Simulation

A simulation runs a model forward to see how a system evolves. Simulations let engineers test a crash in software before building a prototype, let climate scientists compare scenarios that cannot be experimented on, and let biologists observe molecular motion at time scales no microscope can capture.

Pre-processingGeometry, mesh, parameters, boundary conditions
SolveNumerical method advances the model
Post-processingVisualize, compute derived quantities
AssessCheck errors, compare with data

06Optimization

Optimization finds the best design or parameters according to a defined objective, subject to constraints: the lightest bracket that still carries a given load, or the model parameters that best fit observed data. Gradient-based methods follow the slope of the objective downhill; derivative-free and evolutionary methods are used when gradients are unavailable or the objective is noisy.

A key caution: many real objectives have several local minima. An optimizer that reports “converged” may have found a good solution, not the best one.

07High-Performance Computing

High-performance computing (HPC) applies many processors to a single problem. A large simulation domain is split into pieces, each processor computes its piece, and processors exchange boundary data at every step. HPC systems combine multi-core CPUs, accelerators such as GPUs, and fast interconnects.

Two concepts govern HPC performance. Strong scaling asks how much faster a fixed problem runs with more processors; weak scaling asks whether a proportionally larger problem runs in the same time. Amdahl's law shows that the serial fraction of a program limits strong scaling: if 5% of the work cannot be parallelized, the maximum speed-up is 20 times, however many processors are added. Parallel programming is discussed further in cloud and distributed computing.

08Computational Physics

Computational physics simulates physical systems from the subatomic to the cosmological. Examples include molecular dynamics, which integrates Newton's equations for many interacting particles; computational fluid dynamics, which solves the Navier–Stokes equations for liquids and gases; and N-body simulations of gravitational systems. Conservation of energy and momentum is a standard check: a correct simulation of an isolated system should not gain or lose energy beyond its expected numerical error.

09Computational Biology

Computational biology applies models and algorithms to biological data and systems. Sequence alignment uses dynamic programming to compare DNA or protein sequences; population models use differential equations to describe the spread of disease or the growth of species; structural biology uses simulation and learning methods to study how proteins fold and interact. Biological data are often noisy and incomplete, so statistical reasoning is central.

10Engineering Applications

Engineering uses computation throughout the design cycle: finite element analysis for structural stress, fluid simulations for aerodynamics and cooling, circuit simulation for electronics, and digital twins that combine models with live sensor data from an operating system. In regulated fields, simulation results may support certification decisions, which raises the standard of verification required.

11Numerical Accuracy and Validation

Two different questions must be answered before a computational result is trusted:

Verification versus validation
VerificationValidation
QuestionAre we solving the equations correctly?Are we solving the right equations?
MethodConvergence studies, known exact solutions, code testingComparison with independent experimental data
Failure meansA bug or an unsuitable numerical methodThe model's assumptions do not match reality

A convergence study repeats a simulation at successively finer resolution. If the answer stops changing in a predictable way, the discretization error is under control. If it keeps changing, the coarse results cannot be trusted.

12Limitations of Computational Models

  • Model error: the equations omit effects that matter.
  • Parameter uncertainty: input values such as material properties are known only approximately.
  • Numerical error: discretization and rounding introduce approximation.
  • Computational cost: the resolution needed may exceed available resources, forcing simplifications.
  • Chaos: some systems, like weather, amplify tiny differences in initial conditions, limiting how far ahead prediction is possible.

Responsible computational research reports these uncertainties instead of presenting a single number as the answer.

13Frequently Asked Questions

Do I need to be a programmer to study computational science?

You need enough programming to implement and test numerical methods, but mathematics and domain knowledge are equally important. Many people enter the field from physics, engineering or mathematics.

Is a more detailed simulation always better?

No. Extra detail increases cost and can add uncertain parameters. The right model is the simplest one that answers the question with acceptable accuracy.

How is machine learning related to computational science?

Machine learning is increasingly used to build fast surrogate models of expensive simulations and to fit models to data. It must meet the same validation standards. See machine learning applications.

Further reading

  • Heath, M. T. (2018). Scientific Computing: An Introductory Survey (Revised 2nd ed.). Society for Industrial and Applied Mathematics.
  • Trefethen, L. N., & Bau, D. (1997). Numerical Linear Algebra. Society for Industrial and Applied Mathematics.
  • Oberkampf, W. L., & Roy, C. J. (2010). Verification and Validation in Scientific Computing. Cambridge University Press.