01Executive summary
Two milestones whose importance lies not in a breakthrough but in fifty years of accumulated process improvement, and in what a zero fuel cost does to a power system.
Martin Green’s group at the University of New South Wales developed the passivated emitter and rear cell in 1983; it became the mainstream silicon cell architecture decades later. Vindeby in Denmark became the first offshore wind farm in 1991, beginning a scaling programme that has produced the largest rotating machines ever built. Both technologies improved through learning curves rather than discrete invention.
02The learning curve as an engineering phenomenon
Photovoltaic module cost has fallen by a roughly constant proportion for each doubling of cumulative production, sustained over decades. This is a learning rate, and it is a different kind of progress from the one this series has mostly described.
The distinction matters for forecasting. A learning curve is driven by cumulative volume, not by elapsed time. Projections made by extrapolating calendar years have been consistently and badly wrong for photovoltaics, in the direction of pessimism, because deployment volume grew faster than the forecasters assumed. Anyone assessing a technology on such a curve should be projecting against expected cumulative production, not against dates.
Deployment is the research
Improvements come from thousands of incremental process refinements discovered by making things in volume. Subsidising deployment and funding laboratory research are therefore not alternatives; the first generates a class of knowledge the second cannot.
Early cost is a poor predictor
Assessing a technology at low cumulative volume systematically understates it. The question is the slope and how much volume is plausible, not today’s unit cost.
Learning curves are empirical regularities, not laws. They can flatten when a cost floor imposed by materials or energy input is approached, and they can be interrupted by supply constraints in a raw material. A curve fitted over one range does not guarantee behaviour outside it. The record for photovoltaics and lithium cells has been remarkably consistent, but treating that as a guarantee rather than an observation would be an error.
03Photovoltaics: where the losses are
A silicon cell converts photons to charge carriers, and every design improvement addresses one of a short list of loss mechanisms. Understanding the list explains what PERC actually does.
- Optical lossPhotons reflected from the front surface or blocked by contact metallisation never enter the cell. Texturing and anti-reflection coatings address this.
- Transmission lossLong-wavelength photons pass through without being absorbed. A rear reflector sends them back for a second pass.
- Thermalisation lossPhotons more energetic than the band gap deposit the excess as heat. This is inherent to a single-junction cell and is the largest single loss.
- Recombination lossCarriers recombine before reaching a contact, particularly at surfaces where the lattice terminates. Passivation layers reduce this dramatically.
- Resistive lossCurrent flowing through the semiconductor and metallisation dissipates power, trading contact coverage against shading.
The PERC architecture adds a dielectric passivation layer on the rear surface with localised contacts through it. This reduces rear surface recombination and reflects unabsorbed long-wavelength light back into the cell for a second chance at absorption. It addresses two loss mechanisms with one structural change — and, critically, it could be added to existing production lines with a manageable number of extra steps. That manufacturability is why it displaced the previous mainstream design, and why an architecture from 1983 became dominant more than thirty years later.
PERC was not held back by uncertainty about whether it worked. It was held back by cost per watt: the additional process steps were not worth the efficiency gain until other costs had fallen enough to make cell efficiency the dominant term, and until equipment for those steps had matured. A better design waits until the economics of the surrounding system make it worth adopting. Judging a technology as failed because it has not been adopted is a recurring and expensive mistake.
04Wind: why turbines got enormous
Available wind power varies with the swept area of the rotor and with the cube of wind speed. Swept area varies with the square of diameter. These two relationships drive the entire scaling history of the industry.
Doubling rotor diameter quadruples swept area. Height also helps, because wind speed increases with height above ground through the boundary layer, and the cubic relationship means a modest speed increase is a large power increase. Meanwhile the costs that do not scale with rotor size — foundation, grid connection, installation vessel time, consent, maintenance visits — are substantial and are incurred per machine. Larger machines spread those fixed costs over more output.
| Limit | Mechanism | Response |
|---|---|---|
| Blade mass | Naive scaling grows mass with the cube of length while output grows with the square | Carbon fibre spar caps, structural optimisation, slender aerofoils |
| Tip speed | Noise and leading-edge erosion rise steeply with tip speed | Lower rotational speed at larger diameter, erosion-resistant coatings |
| Gearbox reliability | Highly variable torque loading gives poor bearing and gear life | Direct-drive permanent magnet generators, or improved drivetrain load paths |
| Transport and installation | Blades and towers exceed road and crane limits | Segmented blades, offshore siting, purpose-built installation vessels |
| Fatigue loading | Every revolution cycles gravity and wind shear loads for a twenty-year life | Fatigue-driven design, individual pitch control, load-limiting strategies |
The Betz limit deserves a mention because it is frequently misunderstood. No rotor can extract more than about 59 per cent of the kinetic energy in the air passing through it, because extracting all of it would require the air to stop, and stopped air cannot leave to make room for more. Modern rotors achieve a substantial fraction of that theoretical maximum, which means aerodynamic improvement has limited remaining headroom — and is why scaling, reliability and installation cost, rather than rotor efficiency, dominate current engineering effort.
Zero fuel cost means these generators are dispatched whenever available, displacing fuel-burning plant. But output follows weather rather than demand, so the system requires firming — storage, flexible generation, transmission to average across regions, and demand response. Comparing generation technologies on unit cost alone omits this, and comparing them on firmed cost requires stating what portfolio and what reliability standard is assumed. Both numbers are widely quoted and they answer different questions; anyone using either should say which one.
05Takeaways for current practice
- Forecast learning curves against cumulative volume, not calendar time. Date-based extrapolation has failed consistently in this domain.
- Deployment generates knowledge that laboratories cannot. Volume manufacturing is where the incremental improvements are found.
- An unadopted design is not necessarily a failed one. PERC waited thirty years for surrounding economics to make it worth the extra steps.
- Identify which term dominates before optimising. Wind is now limited by fatigue, reliability and installation, not by aerodynamic efficiency.
- State whether a cost figure is firmed or unfirmed. The two answer different questions and are routinely conflated.