Synthetic Data
Synthetic data is artificially generated to resemble real data, used when real examples are scarce, sensitive, or missing important cases.
When real data is not enough
Real data is sometimes too scarce to train on, too sensitive to share, or missing the rare cases a model must handle. Synthetic data is generated artificially to fill these gaps while preserving the statistical properties that make data useful. It is a supplement and a tool, not a replacement for the real data that ultimately defines correctness.
Generation approaches
- Rule-based simulation: encode known mechanics and sample from them
- Statistical sampling: draw from fitted distributions preserving correlations
- Augmentation: transform real examples to create plausible variants
- Generative models: learn a distribution and sample new examples from it
Physics-based synthesis
When the process obeys known laws, a physics-based simulator generates data by modeling the mechanism directly. This is powerful for scientific and engineering domains because the synthetic examples respect the governing equations rather than merely mimicking surface statistics. A fusion program, for instance, can generate large volumes of simulated diagnostic data from physics models to develop and stress-test analysis methods before any hardware exists, with the clear caveat that simulation is not measurement.
The fidelity-privacy-utility triangle
Synthetic data balances three goals in tension. High fidelity means it closely matches the real distribution. Strong privacy means it reveals nothing about any real individual. High utility means a model trained on it performs well on real data. Pushing fidelity too far can leak real records; pushing privacy too far degrades utility. The right point depends on why the synthetic data exists.
Validating synthetic data
Synthetic data must be validated, not assumed good. Distribution comparisons check that marginals and correlations match the real data; a train-on-synthetic, test-on-real evaluation checks that models transfer; and a train-on-real, test-on-real baseline sets the target. The key discipline is that the final measure of any model is always its performance on real held-out data, never on more synthetic data. Overusing synthetic data risks a model that learns the generator's artifacts rather than reality. See data labeling and data-quality validation.