The Climate System from First Principles

Part I — Radiation

1 Planetary energy balance and the greenhouse gap

Start with a planet that has no atmosphere. It intercepts sunlight over a disc of area $\pi R^2$ but radiates from a sphere of area $4\pi R^2$; that ratio, not the fourth power in Stefan–Boltzmann, is where the factor of four comes from. With solar constant $S_0 \approx 1361\ \mathrm{W\,m^{-2}}$ and planetary albedo $\alpha \approx 0.3$,

The observed surface temperature is about 288 K. The 33 K difference is the greenhouse effect, and the whole of radiative climate science is an account of how that gap is set and how it changes. Note what the equation does not say: it says nothing about surface temperature at all. It fixes the temperature of whatever layer does the emitting to space. Everything else is a statement about where that layer sits.

The simplest model that produces a greenhouse effect

Put one layer of atmosphere above the surface. Let it be transparent to sunlight and let it absorb a fraction $\varepsilon$ of the surface's longwave emission, re-emitting $\varepsilon\sigma T_a^4$ both upward and downward (Kirchhoff). The two energy balances are

and eliminating $T_a$ gives

With $\varepsilon = 0$ the surface sits at $T_e = 255$ K; with $\varepsilon = 1$ it reaches $2^{1/4}T_e = 303$ K. The observed 288 K corresponds to $\varepsilon \approx 0.78$. The mechanism is visible in the equation: the atmosphere returns part of what the surface emits, so the surface must emit more than the planet does to push the same $\sigma T_e^4$ out of the top. Stack $N$ opaque layers and the result generalises to $T_s = (N+1)^{1/4}\,T_e$, with each layer a step down in temperature because each must pass the same net flux upward while radiating in both directions.

Use the model below. Drag $\varepsilon$ and watch $T_s$ separate from $T_e$; switch to $N$ layers and watch the profile form. Two things to notice. First, in the leaky single-layer model the atmosphere is always colder than the surface, which is why the "blanket" picture is wrong: the atmosphere does not warm the surface by being warm, it warms it by being between the surface and space. Second, in the $N$-layer model the top layer always sits at $T_e$ whatever $N$ is. Adding layers does not change what space sees; it changes how far below the emitting layer the surface is. That is the emission-height mechanism in its most naked form, and the next chapter puts the layers at real heights.

The model has one dishonesty worth naming before it does damage: emissivity is not a knob. In the real atmosphere $\varepsilon$ is the integral over wavelength of a spectrum full of lines, and adding $\mathrm{CO_2}$ changes it only in the band wings. That is why doubling $\mathrm{CO_2}$ gives $\sim 4\ \mathrm{W\,m^{-2}}$ of forcing, not a doubling of anything. The single-layer model is the right way to see that a partially absorbing layer must warm the surface; Chapter 2 is the right way to see by how much.

Source: Building comprehensive climate science expertise (Dec 2025), including the interactive radiation app.

2 Selective absorption and the emission-height mechanism

Greenhouse gases are not a blanket. $\mathrm{N_2}$ and $\mathrm{O_2}$ are transparent in the infrared because their vibrations produce no change in dipole moment; $\mathrm{CO_2}$, $\mathrm{H_2O}$ and $\mathrm{CH_4}$ have bending and asymmetric stretching modes that do. Absorption is therefore spectrally selective, and the atmosphere has a window near 8–12 µm through which the surface radiates almost directly to space.

Along a path, monochromatic intensity obeys Beer–Lambert, $dI_\lambda = -k_\lambda \rho\, I_\lambda\, ds$, and the cumulative optical depth $\tau_\lambda(z) = \int_z^\infty k_\lambda \rho\, dz'$ measured downward from the top of the atmosphere is the natural vertical coordinate. Adding emission (Kirchhoff: a good absorber is an equally good emitter at the same wavelength) gives the Schwarzschild equation,

whose solution says that radiation escaping to space at a given wavelength originates, on average, from the level where $\tau_\lambda \approx 1$. Adding $\mathrm{CO_2}$ does not widen the atmospheric window; it raises the $\tau = 1$ surface in the band wings. That layer is colder, emits less, the planet runs an energy surplus, and the whole column warms until the new emission level reaches 255 K again. Surface warming is the lapse rate multiplied by the rise in emission height.

The conservation statement $\alpha_\lambda + \varepsilon_\lambda + t_\lambda = 1$ holds per wavelength, not per substance; a cloud can be a near-perfect shortwave reflector and a near-perfect longwave blackbody at once. Shortwave transfer is scattering-dominated (size parameter $x = 2\pi r/\lambda$ sets the regime: Rayleigh for $x \ll 1$, Mie near $x \sim 1$, geometric optics beyond), longwave is absorption-dominated. At the quantum level, absorption requires a photon energy matching a transition between molecular rotational–vibrational states; that is why lines exist, why they broaden with pressure, and why the band wings matter for $\mathrm{CO_2}$ forcing being logarithmic.

Sources: Building comprehensive climate science expertise; Feedback decomposition in climate systems (optical depth, Kirchhoff, photon physics).

3 Feedback decomposition and radiative kernels

Linearise the top-of-atmosphere budget about a reference climate. With forcing $F$ and global-mean surface warming $\Delta T_s$,

The Planck term is the warm-body response and is strongly negative, roughly $-3.2\ \mathrm{W\,m^{-2}\,K^{-1}}$. Water vapour is positive because a warmer atmosphere holds more of it at fixed relative humidity and the extra vapour raises the emission height. Surface albedo is positive. Clouds are the uncertain term and can take either sign depending on altitude: high thin clouds warm by raising emission height more than they reflect, low thick clouds cool by reflecting more than they trap.

A radiative kernel is the partial derivative $\partial R/\partial X$ evaluated once, with a radiation code, for each variable at each level and location. Temperature kernels peak in the upper troposphere because that is where a unit warming most changes emission to space; water vapour kernels are conventionally per unit change in $\ln q$ because the radiative effect of vapour is roughly logarithmic. The kernel separates radiative physics (nearly model-independent) from the climate response pattern $dX/dT_s$ (model-dependent). Differences between models' feedbacks are almost never kernel differences; they are response-pattern differences.

The lapse-rate feedback deserves a critical note that you raised yourself: it is named for a symptom. In the tropics, moist convection ties the upper troposphere to the surface and amplifies warming aloft, which increases emission and gives a negative feedback. In the Arctic, stable stratification traps warming near the surface, so emission aloft barely changes and the feedback is positive. The mechanism is convective coupling versus stratification; "lapse rate" merely names the resulting gradient.

Sources: Feedback decomposition; Radiative kernels; SAI and polar climate (lapse-rate critique).

Part II — The atmosphere

4 Thermodynamics: the engine room

The first law for a parcel, $c_p\,dT - \alpha\,dp = \delta q$, with hydrostatic balance $dp = -\rho g\,dz$, gives the dry adiabatic lapse rate $\Gamma_d = g/c_p \approx 9.8\ \mathrm{K\,km^{-1}}$ immediately. Potential temperature $\theta = T (p_0/p)^{R/c_p}$ is the quantity conserved along dry adiabats, and a profile of $\theta$ is a profile of static stability: $\partial\theta/\partial z > 0$ is stable, and the buoyancy frequency $N^2 = (g/\theta)\,\partial\theta/\partial z$ is the oscillation frequency of a displaced parcel.

Moisture breaks the symmetry. Condensation releases latent heat, so a saturated parcel cools more slowly, $\Gamma_s < \Gamma_d$, and the difference is largest where the air is warm and moist. Equivalent potential temperature $\theta_e$ restores a conserved variable by adding the latent heat that would be released on total condensation. Conditional instability, $\Gamma_s < \Gamma < \Gamma_d$, is the state in which a parcel is stable to dry displacement but unstable once lifted to saturation, which is why convection needs a trigger and why the trigger matters.

Source: Thermodynamics overview (Dec 2025).

5 Convection and precipitation physics

Convective available potential energy is a buoyancy integral between the level of free convection and the level of neutral buoyancy,

and convective inhibition is the same integral below the LFC, where the parcel is negatively buoyant. CAPE is the fuel; CIN is the activation barrier. Severe weather timing is largely a story about when a capping inversion erodes.

Supersaturation alone does not make rain. Köhler theory gives the equilibrium supersaturation over a solution droplet as a competition between the Kelvin curvature term, which raises vapour pressure over small drops, and the Raoult solute term, which lowers it; activation happens when ambient supersaturation exceeds the curve's maximum. Activated droplets grow by diffusion, $r\,dr/dt \propto S - 1$, which is fast for small drops and stalls near 20 µm. Collision–coalescence needs drops large enough to have different fall speeds, so there is a size gap: the warm-rain problem. In mixed-phase clouds the gap is bridged by the Bergeron–Findeisen process, since saturation vapour pressure over ice is lower than over water and ice crystals grow at the expense of droplets.

At climate scale, Clausius–Clapeyron gives $\sim 7\%\ \mathrm{K^{-1}}$ for saturation vapour pressure, extreme precipitation tends to follow it, but mean precipitation is constrained by the atmospheric energy budget to $\sim 2\text{–}3\%\ \mathrm{K^{-1}}$. "Wet gets wetter" is a useful zero-order rule that fails over land.

Sources: Precipitation and climate change; CAPE/CIN; Precipitation physics and cloud microphysics.

6 The boundary layer: K-theory and its failure

The atmospheric boundary layer is the part of the troposphere that feels the surface within an hour. Its physics is turbulent, so the governing quantities are covariances: $\overline{w'\theta'}$, $\overline{u'w'}$, $\overline{w'q'}$. Monin–Obukhov similarity says that in the surface layer these fluxes are constant with height and that dimensionless gradients are universal functions of $z/L$, where the Obukhov length

is the height at which buoyancy and shear production of turbulence are equal. Above the surface layer the convective boundary layer is well mixed and capped by an inversion through which it entrains; at night the stable layer collapses to a few tens of metres and intermittent turbulence makes it the least-well-modelled layer in any climate model.

K-theory, or first-order closure, models a flux as $-K\,\partial\bar\phi/\partial z$, by analogy with molecular diffusion. It works near the surface because eddies there are constrained by the wall. It fails in the mixed layer because coherent structures, thermals, sweeps, ejections and roll vortices, transport heat non-locally: the upper mixed layer has upward heat flux against a slightly stable gradient, which forces $K < 0$. Remedies are counter-gradient terms or explicit mass-flux schemes that represent the thermals as a separate updraught population.

Sources: Atmospheric boundary layer physics overview; Coherent structures and K-theory limitations (Aug 2026).

7 Geopotential, geostrophy and the road to QG

Geopotential $\Phi(z) = \int_0^z g\,dz'$ is gravitational potential energy per unit mass, and geopotential height $Z = \Phi/g_0$ is the coordinate in which parcels respond to energy gradients rather than geometric altitude. Combining hydrostatic balance with the ideal gas law gives the hypsometric equation,

so the thickness between two pressure surfaces is a thermometer for the layer. Warm columns are thick, which is why 500 hPa heights are higher in summer and in the tropics, and why a warm city creates a local thickness anomaly with low-level convergence and upper-level divergence.

On a rotating planet the momentum equation at synoptic scale reduces to a balance between the pressure gradient force and the Coriolis force: $f\,\mathbf{k}\times\mathbf{u}_g = -\nabla_p\Phi$. The wind blows along contours, not across them, and the net force is zero. That last point is the doorway to quasi-geostrophic theory: if the balance were exact nothing would change, so weather is the small ageostrophic residual, of order the Rossby number $Ro = U/fL \sim 0.1$. Expanding in $Ro$ gives the QG vorticity, thermodynamic and potential vorticity equations, and thermal wind $\partial\mathbf{u}_g/\partial\ln p = -(R/f)\,\mathbf{k}\times\nabla_p T$ links the wind shear to horizontal temperature gradients.

Sources: Understanding geopotential; Atmospheric dynamics fundamentals; Climate science tutoring: QG and geostrophy (with force-balance diagram).

8 Heat waves and their compound interaction with cities

A heat wave is a chain: a blocking anticyclone with high potential vorticity anomaly aloft, subsidence warming at the dry adiabatic rate, clear skies, a deep daytime boundary layer that entrains warm air from above, and, over days, the drying of soil that shifts the Bowen ratio $\beta = H/\lambda E$ upward so that more of $R_n$ goes into sensible heat. That last feedback is why the second week of a heat wave is worse than the first even with identical synoptic forcing.

The urban heat island does not simply add to this; it amplifies nonlinearly. Rural surfaces dry and lose their evaporative advantage while the city, already dry, has nothing left to lose, so $\Delta T_{u-r}$ shrinks by day. At night the opposite: the rural stable layer collapses and cools efficiently, the urban layer stays deep and ventilated, and $\Delta T_{u-r}$ re-expands. Air conditioning adds anthropogenic heat with positive feedback on its own demand until saturation. Canyon geometry, via the sky view factor, restricts longwave loss. The physiological limit is a wet-bulb temperature near 35 °C, and humid heat waves reach it at lower dry-bulb temperatures than dry ones.

Sources: Heat wave physics and thermodynamics; Heat waves and urban heat island effects; Urban climate lecture.

9 Land–atmosphere coupling

The surface energy balance $R_n = H + \lambda E + G$ generalises from concrete to every terrestrial surface, but the controls multiply. Soil moisture is the master variable: in the transitional regime between dry and wet it controls $\beta$, and hence the partitioning of $R_n$, and hence boundary-layer growth, and hence cloud and rain, which feed back on soil moisture. Vegetation adds physiology: stomata couple the carbon and water cycles, so rising $\mathrm{CO_2}$ raises water-use efficiency and reduces transpiration even before climate changes. Roughness couples momentum transfer; a forest canopy is a rough, dark, evaporating surface with a displacement height, and the same Monin–Obukhov machinery applies above it.

Land provides the climate system with memory on timescales of weeks to months through soil moisture persistence and of years through vegetation and snow. Spatial heterogeneity in $\beta$ drives mesoscale circulations, and the sign of deforestation's climate effect flips with latitude: tropical clearing warms by suppressing evaporation, boreal clearing cools by exposing snow.

Source: Land-atmosphere interactions overview (Dec 2025).

Part III — The ocean

10 Ocean dynamics foundations

The ocean has absorbed roughly 91 % of the energy imbalance since 1970; it is the climate's flywheel. Its dynamics start from Navier–Stokes in a rotating frame, with the Coriolis term $2\boldsymbol\Omega\times\mathbf{u}$ and, unlike the atmosphere, a nonlinear equation of state: thermal expansion depends on temperature, cabbeling densifies mixtures, thermobaricity makes compressibility depend on temperature.

Scale analysis does the work. Aspect ratio $H/L \sim 10^{-3}$ justifies hydrostatic balance; density variations $\Delta\rho/\rho_0 \sim 10^{-3}$ justify the Boussinesq approximation, in which density appears only in the buoyancy term. Three numbers set the regime: the Rossby number $Ro = U/fL \sim 10^{-3}$ for basin-scale flow (rotation dominates), the Ekman number $Ek = \nu/fH^2 \sim 10^{-13}$ (molecular friction is irrelevant except in thin layers), and the Burger number $Bu = (NH/fL)^2$ (stratification versus rotation, with the deformation radius $NH/f$ as the natural horizontal scale). Geostrophy plus hydrostatics gives thermal wind: horizontal density gradients set vertical shear, so the Gulf Stream's surface velocity can be read from a hydrographic section.

Source: Ocean circulation foundations and dynamics (Jan 2026).

11 The thermocline

Map the ocean column onto the atmospheric boundary layer: mixed layer ↔ convective boundary layer, thermocline ↔ capping inversion, deep water ↔ free troposphere. The thermocline is the band of maximum $N^2 = -(g/\rho_0)\,\partial\rho/\partial z$, and its stability suppresses vertical exchange, which is why the surface ocean can warm while the deep ocean lags by centuries.

Three thermoclines coexist: a diurnal one a few metres deep, a seasonal one that forms in summer and is erased by winter convection, and the permanent thermocline at a few hundred metres to a kilometre. The permanent thermocline's existence is a balance: Munk's abyssal recipe, $w\,\partial T/\partial z = \kappa\,\partial^2 T/\partial z^2$, gives a scale depth $d = \kappa/w$ where slow upwelling of cold deep water is balanced by downward diffusion of heat. At high latitudes salinity, not temperature, sets stratification, and the relevant feature is a halocline. In the equatorial Pacific the thermocline tilts, shallow in the east and deep in the west, and that tilt is the state variable of ENSO.

Source: Thermocline (Jun 2026), including the interactive sounding and ENSO-tilt artefact.

12 ENSO and the Bjerknes feedback

The equatorial Pacific mean state is a coupled asymmetry: trade winds push warm surface water west, the thermocline tilts, the eastern cold tongue upwells cold water, and the west–east SST gradient drives the Walker circulation, a thermally direct zonal overturning that in turn maintains the trades. Bjerknes (1969) recognised that Walker's Southern Oscillation and El Niño were the atmospheric and oceanic faces of one phenomenon.

The Bjerknes feedback is the positive loop: a warm anomaly in the east weakens the SST gradient, weakens the trades, reduces upwelling and deepens the eastern thermocline, which warms the east further. Two sub-feedbacks are distinct: the upwelling feedback (weaker winds, less Ekman divergence) and the thermocline feedback (deeper thermocline, warmer upwelled water). Both are equatorial because $f \to 0$ there: only at the equator can zonal wind stress produce meridional Ekman divergence symmetric about the line, and only there does the thermocline respond directly to the wind.

A purely positive feedback would run away. It does not, and the reason is the subject of Chapter 13: equatorial wave dynamics, the delayed oscillator, and heat-content recharge–discharge.

Source: ENSO coupled instability (Jun 2026).

13 Equatorial waves and the delayed oscillator

Chapter 12 left a loose end that is really the whole problem of ENSO: a positive feedback does not oscillate. The Bjerknes loop, taken alone, predicts a Pacific that departs from its mean state and never returns. An oscillation requires two further ingredients — a sign reversal and a memory — and the ocean supplies both in one mechanism. The memory is wave propagation along the equator, and the sign reversal happens when the wave comes back.

Start from the reduced-gravity shallow-water equations linearised about rest on the equatorial $\beta$-plane, where $f = \beta y$ with $\beta = 2\Omega/a = 2.3\times10^{-11}\ \mathrm{m^{-1}s^{-1}}$:

Eliminating $u$ and $h$ leaves a single equation for the meridional structure of $v$ whose coefficient grows as $\beta^2y^2/c^2$, with $c = \sqrt{g'H}$. It is the quantum harmonic oscillator in disguise: a restoring term proportional to $y^2$, therefore solutions that are Hermite functions confined to a width $L_{\mathrm{eq}} = \sqrt{c/\beta}$, and therefore a discrete ladder of meridional modes rather than a continuum. This is the physical content of “equatorial waveguide”. The equator traps because $f$ vanishes there and grows in both directions, so any parcel pushed off the line is turned back by rotation; the confinement is not a boundary but a restoring force.

Two members of the ladder matter. The Kelvin wave is the degenerate case $v \equiv 0$ everywhere. Zonal momentum and continuity then collapse to $\partial_{tt}u = c^2\partial_{xx}u$ — non-dispersive, speed $c$ — while meridional momentum reduces to exact geostrophy, $\beta y\,u = -g'\partial_y h$. Substituting the eastward-travelling relation $h = (c/g')u$ gives $\partial_y u = -(\beta y/c)\,u$, hence $u \propto \exp(-\beta y^2/2c)$: a Gaussian of width $L_{\mathrm{eq}}$. The westward branch produces $\exp(+\beta y^2/2c)$, which is unbounded and therefore inadmissible. The Kelvin wave travels east only, and that asymmetry is not a modelling choice but a consequence of requiring a finite solution. With $g' \approx 0.02\ \mathrm{m\,s^{-2}}$ and $H \approx 200$ m, $c \approx 2\ \mathrm{m\,s^{-1}}$ and $L_{\mathrm{eq}} \approx 300$ km; the Pacific's 15,000 km takes about three months.

The other member is the long equatorial Rossby wave, which in the low-frequency limit propagates westward at $c/(2n+1)$, so the gravest symmetric mode crawls at $c/3 \approx 0.67\ \mathrm{m\,s^{-1}}$ and needs some nine months to cross. Now the signs. A westerly wind anomaly in the central Pacific launches a downwelling Kelvin wave eastward, deepening the eastern thermocline and warming the cold tongue — the Bjerknes growth of Chapter 12. The same wind stress launches off-equatorial Rossby waves westward carrying the opposite thermocline signal, a shoaling. Those reflect at the western boundary, which is a meridional wall where the only mode able to carry energy back east is the Kelvin wave, and so they return as an upwelling Kelvin wave that shoals the eastern thermocline and cools the cold tongue. The eastern boundary, by contrast, scatters incoming energy mostly into coastal Kelvin waves that leak poleward. The reflection is efficient in the west and leaky in the east, which is why the ocean's memory lives in the western Pacific.

Suarez and Schopf (1988) and Battisti and Hirst (1989) reduced all of this to one line. Writing $T$ for the eastern equatorial SST anomaly,

with $\tau$ the Rossby-westward plus Kelvin-eastward transit time. The linear behaviour follows from $T \propto e^{\sigma t}$, which gives the transcendental dispersion relation $\sigma = a - b\,e^{-\sigma\tau}$. Setting $\sigma = i\omega$ locates the Hopf boundary and separates into

Three conclusions fall out immediately. First, $b > a$ is necessary for a neutral oscillatory mode to exist at all: if the returned feedback is weaker than the local growth, the fastest-growing root of the dispersion relation is real, and the anomaly runs away rather than turning over — a Pacific locked into a warm state rather than one that cycles. Second, $\omega\tau = \arccos(a/b)$, so the period is $P = 2\pi\tau/\arccos(a/b)$, which tends to $4\tau$ in the limit of weak local growth — the delay sets the clock, and the factor of four is the quarter-cycle each leg of the loop occupies. Third, as $a/b \to 1$ the arccosine vanishes and the period diverges: the system is most sluggish and least regular precisely where local growth nearly balances the returned feedback, which is close to where the observed Pacific sits.

The quantitative test is instructive, and subtler than the usual “$4\tau$” slogan suggests. Because $\arccos(a/b) \le \pi/2$, the period is at least $4\tau$ and grows without bound as $a/b \to 1$: the transit time sets a floor on the timescale, not the timescale itself. Taking the wind forcing as centred mid-basin, the Rossby leg is 7,500 km at $c/3$ (about 4.3 months) and the Kelvin return 7,500 km at $c$ (about 1.4 months), so $\tau \approx 5.7$ months and the floor is some 23 months. The observed 2–7 year band then corresponds to $a/b$ between roughly zero and 0.93. That is a real but limited success: the delayed oscillator explains why ENSO lives in the interannual band at all and floors it in the right place, but it cannot pin the period without the feedback ratio supplied from elsewhere. Nonlinearity pulls the other way — saturation lowers the effective growth rate, so a finite-amplitude cycle runs tens of per cent faster than the Hopf estimate, as the figure below shows. What the model cannot produce at all is the irregularity: that needs the seasonal cycle, which phase-locks events to boreal winter by modulating $a$ annually; state-dependent noise from westerly wind bursts; and the nonlinear asymmetry in amplitude and duration between El Niño and La Niña.

Jin (1997) recast the same physics without a lag, by promoting the ocean's memory to a state variable: $h_w$, the thermocline depth anomaly of the western warm pool.

This is an ordinary damped oscillator in two variables, and its eigenvector structure puts $h_w$ a quarter cycle ahead of $T$. The physical narrative is recharge and discharge: under La Niña trades, wind-stress curl drives Sverdrup convergence that recharges warm water volume in the west while the east stays cold; the discharge through an El Niño drains it eastward, leaving the system with depleted heat content and a shoaled warm pool, which is the precondition for the next cold phase. It is the observationally useful form, because $h_w$ is measurable — the warm water volume index from TAO and Argo leads Niño 3.4 by two to three seasons, which a lag model gives no handle on.

One caution, because the correspondence is often stated too casually. Taylor-expanding $T(t-\tau) \approx T - \tau\dot T + \tfrac12\tau^2\ddot T$ in the delay equation does produce a second-order system, and its frequency, $\omega_0 \approx \sqrt{2(b-a)/b}\,/\tau$, agrees with the exact $4\tau$ result to about ten per cent. But the damping coefficient it produces has the wrong sign for all parameters, so the truncated system is unconditionally unstable. The recharge–discharge model is not a truncation of the delayed oscillator; it is a separate derivation in which the wave transit is replaced by an instantaneous Sverdrup balance, legitimate only because the adjustment of the off-equatorial thermocline is fast compared with the oscillation itself.

Written in session, 16 September 2026; closes the question left open at the end of Chapter 12.

Part IV — Cryosphere and carbon

14 The cryosphere as thermal and radiative capacitor

Ice matters through albedo and through latent heat. Ice-albedo feedback is the surface-albedo term of Chapter 3 concentrated at high latitudes, and its strength depends on insolation, which is why Arctic sea-ice loss in summer matters more than Antarctic sea-ice change in winter. Sea-ice growth is a Stefan problem: with conductive heat loss through the ice, thickness grows as $h \propto \sqrt{t}$, so thin ice grows fast and thick ice grows slowly, which is a stabilising feedback on thickness itself.

Ice sheets flow. Glen's law, $\dot\varepsilon = A\,\tau^n$ with $n \approx 3$, makes ice a shear-thinning fluid whose flow is dominated by the basal layers. Where the bed slopes downward inland, marine ice sheet instability follows: grounding-line flux increases steeply with ice thickness, so retreat into deeper water accelerates retreat. Marine ice cliff instability adds mechanical failure of tall exposed cliffs. Mountain glaciers respond on decadal timescales set by their length and mass balance gradient. Permafrost thaw releases carbon that was frozen, not sequestered, and the release is effectively irreversible on human timescales. Freshwater from melt caps the North Atlantic and weakens deep convection, coupling the cryosphere to the thermohaline circulation.

Sources: The cryosphere's role in Earth's systems; Climate science through the Munich lens (Alpine glaciers, föhn).

15 Biogeochemistry and the carbon cycle

Carbon is the master cycle because it couples every reservoir. The ocean takes up $\mathrm{CO_2}$ through three pumps: the solubility pump (cold water dissolves more gas, and it sinks), the biological pump (photosynthesis in the surface, remineralisation at depth), and the carbonate pump (which, counter-intuitively, releases $\mathrm{CO_2}$ because forming $\mathrm{CaCO_3}$ lowers alkalinity). Uptake is limited by carbonate chemistry: the Revelle factor, about 10, means a 10 % rise in atmospheric $\mathrm{CO_2}$ raises dissolved inorganic carbon by only 1 %, and it rises as the ocean acidifies, so the sink weakens.

Land uptake is limited by nutrients. Nitrogen fixation is energetically expensive, phosphorus is supplied only by weathering, and iron limits productivity over a third of the ocean surface. $\mathrm{CO_2}$ fertilisation is real but nutrient-capped. The perturbation response has multiple timescales: decades for the mixed layer, centuries for the deep ocean, millennia for carbonate compensation, hundreds of millennia for silicate weathering. Roughly a fifth of a pulse of fossil carbon remains in the atmosphere after ten thousand years.

Source: Biogeochemistry in climate science (Jan 2026).Phase 2: quantitative carbon cycle

Part V — Detection and intervention

16 Hasselmann: stochastic climate and optimal fingerprinting

Hasselmann's 1976 insight was that a slow system forced by fast white noise integrates it into red noise: the ocean, forced by weather, produces low-frequency variability with no low-frequency forcing at all. That is the null hypothesis against which any detection must work.

Detection is regression in pattern space. Observed change $\mathbf{y}$ is modelled as $\mathbf{y} = \sum_i \beta_i\,\mathbf{x}_i + \boldsymbol\epsilon$, where the $\mathbf{x}_i$ are fingerprints (spatial–temporal response patterns to greenhouse gases, aerosols, solar and volcanic forcing, from models) and $\boldsymbol\epsilon$ is internal variability with covariance $\mathbf{C}$ estimated from control runs. Generalised least squares weights by $\mathbf{C}^{-1}$, which rotates the problem to emphasise directions where the signal is large relative to noise. Stratospheric cooling alongside tropospheric warming is the canonical greenhouse fingerprint because no natural forcing produces that vertical dipole. Practical refinements include regularised covariance estimation and multi-model fingerprints.

Source: Optimal fingerprinting and climate attribution (Dec 2025).

17 Stratospheric aerosol injection as an applied case

SAI raises planetary albedo by placing sulphate aerosol above the tropopause, where residence times are years rather than days. It is included here because it exercises the whole book: it is a shortwave forcing (Chapter 2) applied to a system whose response is set by the feedback structure of Chapter 3, with an Arctic amplification problem driven by the stratification-controlled "lapse rate" term and the ice-albedo term of Chapter 14. High-latitude low-altitude deployment trades a shorter aerosol lifetime for the practicality of using modified commercial aircraft at 13–15 km in subpolar regions, targeting the Arctic directly.

The feedback decomposition of Chapter 3 becomes an examination tool: spatially heterogeneous forcing breaks the global-mean linearisation, cooling may not mirror warming, and kernels assume a stationary base state. Each of those is a question a candidate should be able to answer.

Source: Stratospheric aerosol injection and polar climate impacts (Dec 2025).

Not yet written

Roadmap

Ordered by dependency on what is already here.

• Baroclinic instability and wave–mean-flow interaction: closes Chapter 7 and explains blocking in Chapter 8 properly.
• Ekman dynamics, Sverdrup balance and the thermohaline circulation: the missing middle of Part III.
• ENSO diversity and ENSO under warming: central- versus eastern-Pacific events, and which of $a$, $b$, $\tau$ the mean-state change actually moves.
• Internal variability beyond ENSO: NAO, PDO, AMV, and the stochastic framing of Chapter 16.
• Quantitative carbon cycle: box models, airborne fraction, the impulse response function.
• Paleoclimate: proxies, orbital forcing, the 100 kyr problem.
• Atmospheric chemistry: aerosol physics, stratospheric ozone, short-lived climate forcers.
• Tipping points and bifurcation theory: the formal version of every "runaway" argument above.

Source sessions

Twenty-four teaching sessions between December 2025 and August 2026, plus two planning sessions. Each chapter links to its sources above. Two sessions are not chaptered: the structural overview and prose-craft skill (Dec 2025), which produced the six-pillar map this book follows and the writing method it is written in, and the study plan (Jan 2026), whose Phase 2 is the roadmap above.

Figures so far: the simple climate model (Chapter 1), the emission-height diagram (Chapter 2) and the delayed oscillator (Chapter 13). Other chapters are text-only until asked otherwise.

This first edition was written from session summaries and my own knowledge of the physics. Derivation steps and the specific corrections made in-session (the disc-versus-sphere factor, the per-wavelength conservation point, the lapse-rate renaming) are reproduced from memory of those exchanges; where a chapter needs the full derivation restored from the original thread, ask for that chapter by number.