A Brief History of Empirical Modeling
From Galileo to Machine Learning
Data-Driven Modeling in Science & Engineering
"The most incomprehensible thing about the world is that it is comprehensible. The fact that it is comprehensible is a miracle."
The Renaissance Revolution
- Language opened our eyes to abstract representations
- The Renaissance: space and time can be quantified
- Simple mathematical forms predict behavior with surprising accuracy
Galileo Galilei
- Dropped objects from the Leaning Tower of Pisa
- Deduced relationships between time and displacement
- Newton later formalized: $F = ma$
Key insight: Controllable input measurements linearly affect output behavior
20th Century Paradigm Shifts
- Relativity of time
- Quantum nature of matter
- The advent of computers (most revolutionary!)
Before Computers
Scientists discovered mathematical rules using:
- Pens and paper
- Vivid imagination
- Analytical techniques
Strong bias towards:
- Well-posed linear equations
- Few variables as possible
- Simple building blocks (atoms, cells, agents)
- Systems where superposition applies
The Great Limitation
These assumptions cannot capture the true complexity of:
- Living things
- Ecosystems
- Self-organizing systems
- Turbulent flows
Almost everything that matters!
Computers Change Everything
- Crunch numbers at very high speeds
- Solve complicated nonlinear systems
- Numerical "experiments" / simulations
- Study complex, nonlinear, high-dimensional, multi-scale systems
The Modern Scientific Process
- Derive equations
- Design numerical solvers
- Simulate
- Analyze predictions
- Refine model until observations match
The ML Revolution: Three Factors
1. Moore's Law
Transistors double every 2 years
2. Data Explosion
Data generated online doubles every 2 years
3. Algorithmic Advances
ML algorithms became efficient to train and scale
The Big Question
Can we automate the process of
modeling complex systems?
Empirical Laws
Discovering physical laws =
- Well-designed experiment
- Controlled input causes output to change
- Simple linear fit
- Proportionality constant gets physical meaning
Classical Empirical Laws
- Pascal's Law (1653): $\Delta p = \rho g \Delta h$
- Hooke's Law (1678): $F = -kx$
- Newton's Viscosity (1701): $\tau = \mu \frac{du}{dy}$
- Ohm's Law (1781): $I = V/R$
- Fourier's Law (1822): $q = -k \frac{dT}{dx}$
- Fick's Law (1855): $J = -D \frac{dC}{dx}$
The Ideal Gas Law
$PV = k_B NT$
Combining 2 centuries of linear laws:
- Boyle (1662): $P \propto V$
- Charles (1787): $T \propto V$
- Amonton (1808): $P \propto T$
- Avogadro (1811): $N \propto V$
The Birth of Curve Fitting
- Linear relationships inferred geometrically (with a ruler!)
- Early 19th century: Legendre & Gauss use least-squares
- 1821: Gauss develops full theory of least-squares optimization
The foundation of modern machine learning!
Summary
- From Galileo's experiments to automated discovery
- Computers unlocked nonlinear, complex systems
- ML + Data + Compute = New paradigm
Next: Introduction to Machine Learning