Thermodynamics

Four laws — one of them so obvious it's the zeroth, one of them called "the most profound law of physics" by the people who mean it — that tell you what engines can do, why ice melts, why your laptop needs a fan, and why the universe is slowly running out of usable energy even though total energy is conserved.

1. Why thermodynamics runs every engine on the planet

Thermodynamics is the physics of heat, work, and the invisible constraint called entropy that quietly vetoes half the perpetual-motion machines ever proposed. It was worked out in the nineteenth century by engineers trying to make steam engines burn less coal, and the rules they found turned out to be general enough to describe refrigerators, chemical reactions, neutron stars, black holes, and — more recently — the power budget of a GPU cluster.

Here is the scale of the thing. A 2024 H100 GPU dissipates about 700 watts under load. A rack of 8 of them is 5.6 kilowatts of heat you have to move out of a closet. A 50,000-GPU training cluster dumps 35 megawatts continuously. That heat does not magically vanish; it must be pumped into water or air and then dumped to the atmosphere, and the laws that govern how well that pumping can be done were written down by Rudolf Clausius and Sadi Carnot before anyone had ever seen a transistor. If you work on AI infrastructure, energy, chemistry, climate, or biology, you are doing thermodynamics whether or not the word appears in your job title.

Three reasons to care if you are not a mechanical engineer:

• Infrastructure. Power plants, HVAC, refrigeration, jet engines, rocket nozzles, and data-center cooling are direct applications. Carnot's bound on efficiency is the reason you cannot make a gasoline engine better than about 40 percent without exotic tricks.
• Chemistry and biology. Chemical equilibria, reaction rates, protein folding, and membrane transport are governed by Gibbs and Helmholtz free energies — direct descendants of the second law.
• Information. Shannon entropy in information theory is Boltzmann entropy renamed. Landauer's principle ties the minimum energy to erase a bit to $k_B T \ln 2$. Thermodynamics puts a hard floor under the efficiency of every computation ever performed.

2. Vocabulary cheat sheet

The symbols you will see repeatedly. Glance now; each is defined properly in the sections below.

3. Temperature, heat, and work — three words that are not synonyms

Confusion about thermodynamics almost always starts with the words temperature, heat, and work being used interchangeably in English. In physics they mean three different things, and getting them straight is half the battle.

Temperature is a property of a system in equilibrium. Microscopically it is (up to a factor) the average kinetic energy per degree of freedom. Macroscopically it is what a thermometer measures. The key fact, so obvious it is called the zeroth law, is that if system A is in thermal equilibrium with B, and B is in equilibrium with C, then A is in equilibrium with C. That is what licenses the existence of a temperature scale at all.

Heat is not a substance. It is an amount of energy transferred from one body to another because they were at different temperatures. You cannot "have" heat sitting in a box. You can have internal energy sitting in a box, and some of it may leave as heat if you open the lid to something colder. The distinction matters because "heat" and "work" are the two channels through which energy crosses a system boundary, and the first law depends on knowing which is which.

Work is energy transferred by any other mechanism: a piston pushed, a shaft turned, a charge moved through a potential difference. Mechanical work on a gas in a cylinder is the familiar $P\,dV$:

For a gas expanding at constant pressure the integral collapses to $W = P\Delta V$. For an isothermal expansion of an ideal gas, you need the ideal gas law (coming in section 6) and the result is $W = nRT \ln(V_2/V_1)$, which we will derive in context.

4. The first law — energy bookkeeping, nothing more

The first law of thermodynamics says that energy cannot be created or destroyed; any change in a system's internal energy comes from heat added or work done.

That is it. The first law is a conservation statement, nothing more. What makes it useful is that it lets you track the energy budget of any process, no matter how complicated, as long as you are careful about which flows are heat and which are work.

A few named processes fall out of the first law by holding one variable fixed:

• Isothermal (constant $T$). For an ideal gas, $U$ depends only on $T$, so $\Delta U = 0$, and therefore $Q = W$: every bit of heat added is converted directly to work done by the gas.
• Adiabatic (no heat exchanged, $Q = 0$). Then $\Delta U = -W$: the gas cools when it expands because the work comes out of its internal energy. This is why a can of compressed air gets cold when you discharge it.
• Isobaric (constant $P$). Work is $P\Delta V$, heat is $nC_P \Delta T$, and internal-energy change is $nC_V \Delta T$, where $C_P$ and $C_V$ are molar heat capacities.
• Isochoric (constant $V$). No work, so $\Delta U = Q$ and all added heat goes into internal energy.

The key relation between the two heat capacities of an ideal gas is:

5. The second law — the rule that kills perpetual motion

The first law forbids getting energy out of nothing. The second law forbids something subtler: getting work out of energy that is already there, unless some of it is thrown away as waste heat. Clausius's original 1850 statement:

Kelvin gave an equivalent version focused on engines:

These two statements look different and sound like folk wisdom. Clausius showed they are equivalent and extracted from them a new state function, entropy, defined for a reversible process by:

The second law then states, in its cleanest form:

A consequence: when heat $Q$ flows from a hot reservoir at $T_H$ to a cold one at $T_C$, the hot side loses entropy $Q/T_H$ and the cold side gains $Q/T_C$. Since $T_H > T_C$, the cold gain is larger, and total entropy grows. Reversing that flow would shrink total entropy — forbidden.

6. The ideal gas law and $PV$ diagrams

The simplest system you can do real calculations with is the ideal gas: a dilute collection of non-interacting particles. The equation of state that ties its four variables together is:

The microscopic derivation, which we do carefully in section 8, shows that temperature in an ideal gas is literally a measure of average kinetic energy: $\frac{1}{2} m \langle v^2 \rangle = \frac{3}{2} k_B T$.

A PV diagram plots pressure on the vertical axis against volume on the horizontal axis. Each point is a state. A curve is a process. The area under a curve is $\int P\,dV$, the work done by the gas. A closed loop is a cycle, and the area it encloses is the net work done per cycle — the basis of every heat engine on the planet.

An isothermal process at temperature $T$ traces out a hyperbola $P = nRT/V$. An adiabatic process for an ideal gas traces out a steeper curve $PV^\gamma = \text{const}$, where $\gamma = C_P/C_V$ is the heat-capacity ratio ($\gamma \approx 1.4$ for air at room temperature). A simple isothermal expansion from $V_1$ to $V_2$ gives work:

7. Heat engines and the Carnot bound

A heat engine is any device that takes heat $Q_H$ from a hot reservoir at $T_H$, converts some of it to work $W$, and dumps the rest $Q_C$ into a cold reservoir at $T_C$. Its efficiency is the fraction of input heat that turns into useful work:

In 1824, twenty-five years before the first law was written down, Sadi Carnot asked a sharp question: What is the best efficiency possible for a given $T_H$ and $T_C$? His answer, now called Carnot's theorem, is that the maximum is achieved by a reversible cycle of two isothermal steps (at $T_H$ and $T_C$) and two adiabatic steps connecting them:

Worked example — the coal plant. A 1 GW electric output power plant running at 40 percent thermal efficiency must burn $1/0.40 = 2.5$ gigawatts of coal continuously, which means it must dump $2.5 - 1.0 = 1.5$ gigawatts of waste heat into a river or cooling tower. At the Carnot limit of 62.5 percent it would still dump $0.6$ gigawatts. You cannot design that waste heat away; it is required by the second law. This is also why combined cycle plants (gas turbine feeding a steam turbine) beat 60 percent efficiency: the exhaust of the gas turbine, at around 900 K, is still hot enough to usefully drive a second Carnot-like cycle down to 300 K.

Refrigerators and heat pumps are engines run backward: you do work $W$ to pump heat $Q_C$ out of a cold space into a hot one, moving $Q_H = Q_C + W$ total. Their figure of merit is the coefficient of performance, $\text{COP}_{\text{ref}} = Q_C/W$, with a Carnot-style upper bound of $T_C/(T_H - T_C)$. A home freezer with $T_C = 260\,\text{K}$ and $T_H = 300\,\text{K}$ has a theoretical maximum COP of $260/40 = 6.5$ — each joule of electricity could in principle move $6.5$ joules of heat out. Real freezers hit about half that.

8. Statistical mechanics — why any of this works

Thermodynamics, as we have presented it, is a phenomenological theory. It tells you what is allowed and what is forbidden, but it does not explain where the rules come from. Statistical mechanics is the bridge from microscopic physics to those rules. It was Boltzmann's contribution, and it is the reason entropy can be defined without waving your hands at "reversibility."

The central idea is simple. A macroscopic state (say: a mole of gas at temperature $T$, volume $V$, pressure $P$) corresponds to an astronomical number of compatible microscopic states — individual positions and velocities of all $10^{23}$ molecules. Call the number of microstates consistent with a given macrostate $\Omega$. Then:

This formula is engraved on Boltzmann's tombstone in Vienna. It is the most succinct statement of the microscopic origin of the second law. Entropy increases because the system wanders through state space and state space is overwhelmingly dominated by high-entropy configurations.

From here you can derive the Boltzmann distribution. A small system in contact with a large heat reservoir at temperature $T$ has probability of being in a given microstate $i$ (with energy $E_i$) equal to:

The partition function $Z$ is the magic object. Once you have it you can get everything:

For the ideal gas, the single-particle partition function gives $\langle E \rangle = \frac{3}{2} k_B T$ per particle — the equipartition theorem in its simplest form. Multiply by $N$ and differentiate: you recover $PV = Nk_B T = nRT$. Thermodynamics was right, and we now know why.

One more classic application: the Maxwell-Boltzmann speed distribution. The probability density that a molecule in an ideal gas has speed $v$ is:

9. Interactive: the Maxwell-Boltzmann distribution

Pull the temperature slider below and watch how the speed distribution of nitrogen molecules (the dominant component of air) changes. Notice two things. First, the peak shifts to the right as $T$ grows — hotter gas has a faster "most probable" speed. Second, the whole curve gets wider and shorter, preserving total area. The high-speed tail, where interesting chemistry and escape physics live, grows much faster than the peak does. A small temperature bump produces a dramatic increase in the fraction of molecules above any fixed "activation" threshold.

10. Free energy and chemical potential — when $T$ and $P$ are fixed

Most real-world thermodynamics does not happen in closed boxes at fixed energy. It happens in beakers at fixed temperature and pressure, or in cells at fixed temperature and chemical environment. In those settings, the quantity that tells you what a system can actually do is not internal energy but free energy.

The Helmholtz free energy $F$ is relevant at fixed $T$ and $V$:

The Gibbs free energy $G$ is relevant at fixed $T$ and $P$ — the natural setting for chemistry, biology, and anything open to the atmosphere:

For systems that can exchange particles — chemical reactions, phase transitions, electrochemical cells, the interior of a star — you also need the chemical potential:

Equilibrium conditions fall out immediately. Thermal equilibrium: $T$ equal throughout. Mechanical equilibrium: $P$ equal throughout. Chemical equilibrium: $\mu$ equal throughout for every species that can move. These three conditions, along with the second law, determine every phase diagram, every chemical equilibrium constant, and every osmotic pressure in biology.

11. Simulation in code

Two small programs. The first samples the Maxwell-Boltzmann distribution for nitrogen and reports the mean speed and rms speed, checking them against theory. The second runs a toy Carnot cycle on an ideal gas and prints the efficiency.

import numpy as np

# Physical constants
# ----------------
kB = 1.380649e-23   # J/K
amu = 1.66054e-27   # kg
m_N2 = 28.0 * amu   # nitrogen molecule mass

T = 300.0           # kelvin — room temperature
N = 200000          # number of molecules

# Sample each Cartesian velocity component from a zero-mean Gaussian
# of width sqrt(kB T / m). This is the Maxwell-Boltzmann distribution
# factored into three independent 1D Gaussians.
sigma = np.sqrt(kB * T / m_N2)
vx = np.random.normal(0.0, sigma, N)
vy = np.random.normal(0.0, sigma, N)
vz = np.random.normal(0.0, sigma, N)

# Speed is the magnitude of the velocity vector
v = np.sqrt(vx**2 + vy**2 + vz**2)

v_mean_theory = np.sqrt(8 * kB * T / (np.pi * m_N2))
v_rms_theory  = np.sqrt(3 * kB * T / m_N2)
v_peak_theory = np.sqrt(2 * kB * T / m_N2)

print(f"mean speed  (measured): {v.mean():8.2f} m/s")
print(f"mean speed  (theory):   {v_mean_theory:8.2f} m/s")
print(f"rms speed   (measured): {np.sqrt((v**2).mean()):8.2f} m/s")
print(f"rms speed   (theory):   {v_rms_theory:8.2f} m/s")
print(f"most-probable (theory): {v_peak_theory:8.2f} m/s")

# For N2 at 300 K: mean ~ 476 m/s, rms ~ 517 m/s, peak ~ 422 m/s.
# These numbers are faster than the speed of sound in air (~343 m/s)
# — the sound speed is smaller because it is an average with a factor of gamma in it.

import numpy as np

# Ideal-gas Carnot cycle
# ----------------------
# 1 mole of ideal gas, gamma = 5/3 (monatomic)
R = 8.314        # J/(mol K)
n = 1.0          # mole
gamma = 5.0 / 3.0

T_H = 800.0      # hot reservoir, K
T_C = 300.0      # cold reservoir, K

# Stage A: isothermal expansion at T_H from V1 to V2
V1 = 1.0e-3      # m^3 (1 liter)
V2 = 3.0e-3      # m^3 (3 liters)
Q_H = n * R * T_H * np.log(V2 / V1)   # heat absorbed
W_A = Q_H                                  # U unchanged at constant T

# Stage B: adiabatic expansion from T_H, V2 to T_C, V3
# For an adiabat: T V^(gamma-1) = const
V3 = V2 * (T_H / T_C) ** (1.0 / (gamma - 1.0))
W_B = n * R * (T_H - T_C) / (gamma - 1.0)

# Stage C: isothermal compression at T_C from V3 to V4
# V4 chosen so stage D closes the loop back to (T_H, V1)
V4 = V1 * (T_H / T_C) ** (1.0 / (gamma - 1.0))
Q_C = n * R * T_C * np.log(V4 / V3)   # negative — heat rejected
W_C = Q_C

# Stage D: adiabatic compression from T_C, V4 to T_H, V1
W_D = -W_B   # adiabats are symmetric in sign

W_net = W_A + W_B + W_C + W_D
eta   = W_net / Q_H
eta_carnot = 1.0 - T_C / T_H

print(f"Q_H = {Q_H:.1f} J   (heat in from hot reservoir)")
print(f"Q_C = {Q_C:.1f} J   (heat out to cold reservoir, negative)")
print(f"W_net = {W_net:.1f} J   (work out, = |Q_H| - |Q_C|)")
print(f"eta (cycle)  = {eta:.4f}")
print(f"eta (Carnot) = {eta_carnot:.4f}")

# Output:
#   eta (cycle)  = 0.6250
#   eta (Carnot) = 0.6250
# A real Carnot cycle saturates the bound exactly — that's the whole point.

12. Cheat sheet and see also

See also

Further reading