Reducing the Cost of Solar Energy

My previous post discussed the erroneous reasoning that has led some commentators to believe that there’s a Moore’s Law for solar; there isn’t.  Electrons are a massive commodity industry, and the only way solar energy will gain any significant market share is if becomes cost-competitive with natural gas and coal power plants, and is able to scale into the hundreds of gigawatts.  Clay Christensen echoes some of these sentiments in his article here.  Let’s see what the prospects of that happening by the end of the decade are.

I will argue that no new science is required for this to happen.  Science is hard and costly – as the Solasta story attests.  We know how to make tiny, high-efficiency thin-film solar cells in the lab from compound semiconductors like CIGS and CdTe, but manufacturing them reliably requires years of tweaking and optimization.  The two most promising CIGS companies, Nanosolar and Miasole, have had disastrous results.  After a decade of work and over a billion dollars in funding, Nanosolar’s latest valuation is a paltry $50 million and Miasole was just sold for $30 million to a Chinese firm.  Forget about third-generation quantum dots, polymer solar cells, nanowires (I’ve heard they’ve had reproducibility issues), and “pyro-nano-quantum-thingamajig“s.  Even if we had a breakthrough, solar investors are risk-averse and will want to see 20 years of field data before any technology becomes truly “bankable.”  Furthermore, new manufacturing technologies are just not cost-competitive against silicon, which has a massive scale advantage.  We already know how to mass-produce ~20% efficient silicon solar cells and assemble them into modules at a price of 70 cents/Watt-peak.  Silicon is probably the single most-studied element and material in the world, thanks to its half-century history role at the center of the trillion-dollar semiconductor industry.  Not only is it a single, well-understood element (compared to a crazy phase diagram for CIGS), it’s one of the most common elements in the earth, and we’re already refining over 200,000 TONNES of it annually.

What is needed is engineering to improve the cost of traditional silicon solar technology by a factor of 2 – that’s it.  Let’s begin by analyzing some up-and-coming technologies that can reduce the cost of silicon solar cells, and then look at balance-of-system costs to see whether we can achieve this target.  All of the approaches mentioned here involve ways of reducing the kerf losses of the silicon wafer.  The kerf is the “sawdust” created when bricks of silicon are put through a wire saw to cut them into thin wafers of about 180 micrometers in thickness.  Because the wire itself is 120 microns thick, and the sawing results in micro-cracks in a layer of the wafer 10 microns thick on either side, you end up with losses of 120 + 20 = 140 microns, or close to (140/(140+180)) = 44%.  The industry is moving to thinner wires coated with diamond abrasives, but as you make the wafer thinner you enter what I call the silicon wafer valley of death – a wafer with a thickness in the range of 160-60 microns is just to brittle to survive a typical fab process. At around 50 microns, the wafers are flexible.

A number of companies are pursuing kerfless wafers.  1366 Technologies, spun out of MIT, is one of the most interesting.  Their “direct wafer” technology apparently uses just 3 grams of silicon per watt of output, compared to the current industry best, which is 5.3 grams/watt for SunPower’s cells (down from 13 grams/watt in 2004).  Based on my estimates, though, reducing the silicon usage by half will reduce the module cost by 25%, and the entire system (and thus the LCOE) by only 12.5% (as each is typically half of the cost of the higher level system).  only results in a previous ribbon solar throughput.  Their capital expenditure looks good, though at about 33 cents/watt of capacity.  Seeing as how Ely Sachs previously invented the string ribbon technology that never took off (likely due to low throughput), he’s hopefully taken the lessons learned from that and applied it to the new design.  So how does it work?  Despite multiple press stories where we learn that “company officials will not say just how they do that“, investigative journalism doesn’t seem to be the forte of the New York Times.  For if they had bothered to do a bit of research, they would have uncovered the masters thesis of one of the company’s employees, along with two patents (here and here) that go into great detail about the process.  In the first step, a square-shaped chuck makes contact with a melt of silicon, and pulls away after a thin wafer has been formed on its surface.  To remove the wafer from the chuck, the chuck has tiny holes in it through which a pressurized gas is blown to release the wafer.  Because the wafer was crystallized so quickly, however, it’s got extremely poor quality.  In order to improve its quality and grain size, its recrystallized by encapsulating it in a thin layer of silicon dioxide or silicon nitride, and then sent through a furnace which re-melts the silicon and cools it down slowly.  The encapsulant is then chemically removed, and voila – a kerfless silicon wafer!  How they’ve managed to get carrier lifetimes of 20 microseconds from such a process where impurities at the interface would be contaminating the crystal is quite impressive.  Oxidizing silicon wafers typically is not a very cheap process, as it requires high temperatures and quite a bit of time, but using silicon nitride deposited through CVD could be a cheaper route.  Here’s a prototype of the recrystallization furnace from Eerik’s thesis:

The other really fascinating approach is that of exfoliation, which is being pursued by Twin Creeks Technologies (whose gigantic Hyperion machine is pictured below) as well as Silicon Genesis.  The wafers are placed in a vacuum where they are implanted with high-energy protons which induce cleavage of the wafers.  Not only have they managed to create a tool with a throughput of 6 MW/year, but they’ve figured out how to support a 20-micron thin wafer so it’s compatible with existing production lines and do light-trapping on it as well.  If the expensive tool, which promises productivity gains of 12X, gets orders from large Chinese producers, this could be the future of the industry.

IBM and Astrowattare trying to cleave wafers as well, but are doing it through a wet chemistry approach that IBM calls spalling.  They both seem to be doing electroless nickel plating, followed by an annealing step to incorporate hydrogen ions into the lattice, followed by a lift-off of the nickel layer (along with a thin layer of silicon).

A few other approaches are in the very early research stages, but are nonetheless interesting.  Solexel has made some progress with porous silicon technology licensed from IMEC, which grows crystalline silicon from silane gas through CVD.  Crystal Solar is trying a similar approach on etching silicon wafers to create pores, then doing epitaxy, depositing a ceramic slurry, and then exfoliating a thin layer – efficiencies are still low, though.  Others include a liquid phase epitaxy approach being pursued by Ribbon Technology, a pretty unique continuous sheet crystallization method called horizontal ribbon growth underway at Varian, and a method analogous to float glass production where the silicon floats atop a layer of molten tin at CMU.

---

On the manufacturing side, there are a number of existing techniques that could be fine-tuned and developed to make incremental improvements in cost.  These include replacing the costly silver paste used for metallization on the front of the cell, shifting production to “quasi-mono” ingots with higher efficiency, and automating the tabbing and stringing of solar cells.

The solar industry consumes a lot of silver – almost 11% of the world’s supply.  It’s become extremely expensive recently, “adding $23.52 to the cost of each panel,” based on some estimates.  There have been many attempts to replace silver with copper, typically by using a thin layer of electroless nickel as a diffusion barrier, with copper electroplated on top through a self-alignment process.  Due to the low throughput of the process compared to screen printing silver paste, I believe SunPower is the only company that has adopted this so far.

It was surprising to learn that many of the largest Chinese manufacturers still have not automated the stringing of individual cells in series.  Suntech must not have showed this part of the factory to Steven Chu, who said during a tour that “It’s a high-tech, automated factory. It’s not succeeding because of cheap labor.”  Suntech claims that the machines break more of the fragile solar cells than humans do, but how can this process ever scale?  Hopefully with metal-wrap-through solar cells where all the contacts are on the back, some manufacturers will be able to automate this process without breakage.

(courtesy Technology Review)

The third technology that is starting to be adopted is directional solidification of ingots, known colloquially as “quasi-mono” because of its larger grain sizes than multicrystalline ingots.  This results in higher efficiency, but at a lower cost than if one had used CZ crystal pullers to make a single crystal.  Also known as vertical gradient freeze, the technique extracts heat from the crucible so that the silicon freezes over time in one direction, at a rate of about 1-2 cm/hour.  Some of these furnaces (from the likes of GT, AMG, ALD, PVA, ECM, Centrotherm, JYT, Jing Gong Technology, and JingSheng) can take in a charge of 650 kilograms at a time and achieve a throughput of 10 MW a year. The cost breakdown of one of these furnaces per run is about $176 for 4,000 kWh of electricity, $27.5 for 60 cubic meters of natural gas, $1,248 for a fused silica crucible (such as those made by Ceradyne, Vesuvius, Sinoma, and Saint Gobain), and $700 for 20 hours of labor.  In the end you end up with quality close to that of CZ ingots ($24.5/kg) at the price of multicrystalline ingots ($10/kg).

If you look at the non-cell costs of a solar module and normalize them by area, you end up with a breakdown like this, totaling $31/square meter (left graph).  Including the solar cell, you have a breakdown of costs like so (data is from 2012, right graph):

---

Finally, let’s take a look at balance-of-system (BOS) costs, which include the inverter, racking and mounting hardware, installation labor, and permitting.  It’s clear that the industry has reached an inflection point, where the cost of the solar module itself is now less than 50% of the total installed cost of a system.  For example, First Solar estimates that in 2014 its module cost will be $.63 and BOS costs at $.98/watt.  Clearly there is more fat to be trimmed from the BOS side than the module itself, especially for smaller residential and commercial systems.  It’s certainly possible, as permitting costs for installations in Germany at less than 1/10th of what they are in the U.S.  While companies like Clean Power Financeare working on streamlining the red tape around structural permits in different counties, it would be much more helpful if states simply removed the need for permits and inspections for small installations, like the Vermont legislature has done.As an investor at NEA (investors in Qbotix, formerly known as Black Swan Solar) statedearlier this year:
“

If you look at solar, cell prices are plummeting upstream. That means the cost bottleneck is going to shift into the balance of system, not the [solar] cell and module, but all the other stuff: the tracker, the inverter, etc. That has been a real issue. They haven’t materially gone down. [Because] the cell and module costs have plummeted, a much bigger percentage cost is now balance of system costs

”
The DOE’s Sunshot program has espoused the “plug and play“ paradigm, as has the Rocky Mountain Institute’s study – but I believe more work is needed on lower cost racking and installation.  Installing panels one at a time on solar farms is difficult to scale – a tiny solar farm of 45 MW in capacity requires almost half a millionpanels to be installed over an area of 500 acres.  By hand. I’d also like to point out that the oft-cited graph of when we can expect grid parity in different countries is over-simplified because it does not account for the differing costs of installation labor in those countries.

---

Ultimately, the most significant impact on the adoption of solar may not come from technology at all, but new financing models such as securitization by banks and startups like Solar Mosaic.  Even if we get to $1/watt installed costs, the world is not going to switch from coal to solar overnight – it’ll still take decades, as Vaclav Smil has pointed out.  As with any energy technology, the factors limiting the speed of growth will be the cost of capital and the time involved in working though a thicket of regulations around power lines and land use.  As Dick Swanson has mentioned, “For the large-scale systems, you can’t just snap your fingers and install hundreds of megawatts overnight.”  It doesn’t sound like solar is going to be able to scale to meet the challenges of global warming anytime soon.

Henry Samueli, Co-founder of Broadcom

Henry Samueli (otherwise known as the ‘good Henry‘ to Broadcom investors) recently gave a talk about the origins of Broadcom and his research.

or download the MP3 here.

Some excerpts:

• He worked at TRW (a defense contractor) on military broadband radios for 5 years after getting his PhD on digital signal processing from UCLA in 1980.
• He believes it would be hard to replicate the success of Broadcom today, as at the time it was founded in 1991 they had luck and timing on their side – they were at
  the right place at right time, with the right technology to address market need (broadband communications)
• They now have 5,000 employees in California and about the same number overseas in 60 different R&D locations around the globe
• Broadcom sells 2 BILLION chips a year – half of which is for wireless
• The company was started with two $5,000 checks, and they never took VC funding
• He started teaching some courses at UCLA and published his research on chips for broadband communications
• Once companies at conferences started asking about the chips, they got funding from DARPA for the work and got their first contracts from Scientific Atlanta (for cable TV boxes) and Intel (for Ethernet transceivers)
• When their share price was high ($200), they used it like Monopoly money to gobble up startups that were the best in the world in some area that they had a gap in
• They created 2 classes of stock, class A for the open market, and class B for employees which held 10 votes per share.  The founders were able to retain control of the company because they had >50% of the voting power even though their ownership amounted to 5%
• As a startup, he spent most of his time on recruiting, and kept the ratio of engineers to non-engineers high.  He was working 18 hours a day, 7 days a week at the start
• The advertised wireless data rates of cell phone carriers are theoretical best-case scenarios, if you happen to be the only user standing 10 feet away from the cell tower
• 20% of the work leads to 80% of the progress
• 2G, 3G, and 4G have been made possible by Moore’s law, but it won’t last forever – maybe 10 years from now we’ll start to level off
• His PhD training in solving and methodologically approaching problems was invaluable, more so than the actual project

There’s been some talk lately, at least in the US about the spectrum crunch becoming a huge bottleneck to the mobile industry.  Steve Perelman has been working on a radio technology that claims to overcome the Shannon limit for channel capacity; from reading the patent it appears that the main advantage compared to MIMO antennas is that the system only requires a single antenna on the mobile.  It sounds very similar to network MIMO and recent work that’s been done on interference alignment algorithms and dirty paper coding.

My first introduction to information theory came by way of Professor Thomas Cover who was teaching a course on game theory at the time.  He was a wonderful teacher – “the jewel in Stanford’s crown” – and I’m just beginning to scratch the surface of some of his work.  Thank you, Prof. Cover.

the Mythical Moore’s Law for Solar Energy

It’s understandable when people without any scientific or industry experience to talk about a Moore’s Law for solar energy, but those who understand the technology should know better.  The reason it’s so easy to fall into the trap of believing this is due to the superficial similarities between solar cells and transistors; they’re both typically made from silicon wafers, and utilize many of the same processes during manufacturing (diffusion furnaces/ion implantation, thin film deposition, etching, plating, etc).

And so when people hear that the prices of solar cells are dropping, and they’re made from silicon, it seems that they readily assume that it’s Moore’s Law at work.  Let’s see what Al Gore said a few years ago, as quoted in Vacliv Smil’s book Energy Myths and Realities:
“

[the] price of specialized silicon used to make solar cells was recently as high as $300/kg.  But the newest contracts have prices as low as $50/kg.  You know the same thing happened with computer chips – also made out of silicon.  The price paid for the same performance came down 50% every 18 months – year after year.

”
More or less the same argument was made by Paul Krugman in this New York Times op-ed, in a Scientific American article here, and in Kevin Kelly’s book What Technology Wants.

What’s surprising is that you have extremely well-qualified people making similar proclamations, though without the fallacious reasoning.  Steven Chu, for example, is a Nobel laureate in physics and current head of the Department of Energy.  As stated in this 2004 conference paper, he conflates learning curves with Moore’s Law (and also fails to distinguish between windmills – which provide mechanical power, and wind turbines – which provide electrical power):
“

Every technology seems to follow a Moore’s Law curve, which means that the cost effectiveness improves exponentially as a function of the overall money invested in the deployment of that technology. Figure 1 shows Moore’s Law curves for photovoltaics, windmills, and gas turbines. As you put more money into a technology, that drives the price down.

”
Next we have two CEOs of solar companies.  Let’s start with this statement from the head of NRG:
“

A form of Moore’s law — the doubling every two years of the number of transistors that can be placed on an integrated circuit — applies to photovoltaic technology, according to Crane. In the last two years, the delivered cost of energy from PV was cut in half, he said.  NRG expects the cost to fall in half again in the next two years, which would make solar power less expensive than retail electricity in roughly 20 states, he said.

”
A few months ago the founder of Suntech, the world’s largest manufacturer of solar panels, gave a talk at Stanford.  One of his slides compared the falling costs of solar to those of digital cameras, cell phones, and DVD players soon after they were first introduced.  He thankfully didn’t explicitly say anything about Moore’s Law, but the comparison nonetheless is not quite apples-to-apples.

So what’s the issue with all of these kinds of statements?

• Moore’s Law is specific to the number of transistors on an integrated circuit, and is not applicable to other fields just because they bear some superficial resemblance with the chip industry.
• The rate of progress in the solar industry (~30X cost reduction in the past three decades) is orders of magnitude below the ~40,000X increase in the number of transistors of a microprocessor of today compared to ones like the Intel 186 from 1982.
• The physics that limit and constrain solar cells are different from those of processors.  With solar cells, you’re fighting thermodynamics, and you’re never going to win the game.  Although we’ll probably get to a 50% efficient multi-junction solar cell soon, you can’t go past 100% – the ceiling is fixed.  With computation, though, there’s still plenty of room at the bottom.  Whether its single-electron transistors or room temperature quantum computers, we’re still very far away from nearing the ultimate limits of computation.
• The generally accepted formulation of the solar industry’s learning curve, which states that prices of solar modules will fall 19% every time production volumes double, makes no predictions about how quickly volumes will double (it may take one year or a hundred).  In contrast, Moore’s Law has a fixed time frame for each doubling of the number of transistors (18 months in its most recent, amended form).
• The market for solar panels is somewhat elastic; prices for solar panels have dropped recently not because of some major technological or manufacturing advance, but because supply has continued to increase, while demand has decreased (likely due to reductions of feed-in tariffs in Europe).  Solar panels are essentially commodity items (albeit, relatively expensive ones) that produce electrons, and any commodity’s price tends to be dictated by the pull between supply and demand, like a feedback loop. As the industry continues down its learning curve, however, the manufacturing cost/Watt-peak of the panels will decrease, putting downward pressure on the prices, irrespective of supply and demand.  You can think of the learning curve as a trend line with a negative slope, around which the market forces oscillate.  Five years ago, when polysilicon was in short supply, the price shot up above the trend line, and today when demand has dropped, we’re below that trend line. The amplitude of the oscillations tends to be large because while demand can change fairly quickly through changes in legislation (feed-in tariffs, tax incentives) and the economy (financial crises that suck up liquidity), it takes time for supply to adjust (polysilicon plants cost over a billion dollars and take years to build).
• The cost structures of solar cells are very different from those of microelectronics.  Silicon wafers make up about half the cost of solar cells; but when you can fit thousands of microchips onto a single wafer, the costs of silicon raw material per chip are minuscule.  The real costs lie in the $6 billion or so it takes to setup a new fab (expensive UV lithography equipment) and R&D.
• The areas where we have seen really stunning improvements over time have all been able to take advantage of engineering on the micro or nano-scale.  These include CPUs (Moore’s Law), hard disk drives (Kryder’s law), fiber optic communications (Butter’s law), LEDs (Haitz’s law), and DNA sequencing.  Solar cells and LCDs, however, have improved much more slowly, even though they too have borrowed many tricks from the microelectronics/optoelectronics industry.  The underlying reason is that the solar and LCD industries are building on the macro scale.  Their key metric is cost/area, whether it’s a solar module or a television.  Prices for LCDs have certainly come down, but not by the orders of magnitude we’ve come to expect (courtesy Hendy Consulting):

• The fundamental disparity between the technology learning curves can perhaps be understood by characterizing them based on density – of power or information flow.  A modern CPU can easily generate fluxes of over 300 Watts/square centimeter:
• Or consider high-brightness LEDs, dissipating heat at over 100 W/square centimeter.  Single mode fiber optic cables can transmit tens of gigabits per second along a fiber just a few micrometers in diameter.  Fourth generation DNA sequencers will use nanopores to read individual DNA bases much faster and more accurately than current technologies.  Hard drives have reached areal densities which are staggering (courtesy IBM):

• Contrast this with solar cells that receive no more than 1,000 Watts per square meter of area, which amounts to a flux of just 0.1 Watt per square centimeter.  Similarly, battery performance improvements in terms of energy density have been moving at a snail’s pace for the past century (courtesy this paper).  Products made on the micro and nano scale can pack a lot of power and information into a tiny space, allowing millions of units to be mass-produced with negligible raw material costs.  Products engineered on the macro scale have a much harder time improving performance or manufacturing.

In summary:

1. Just because a technology has a learning curve, doesn’t mean it’s Moore’s Law.
2. Not all exponentials are equal.
3. Macro-scale engineered solar panels and batteries improve performance over time, but nowhere near as fast as compared to micro-scale engineered products.

the Diffusion of Innovations, or Lack Thereof

Everett Rogers wrote the classic work on the difficulties of spreading new ideas and solutions, none more elucidating than the case of the British Navy and the cure to scurvy:
“

In the early days of long sea voyages, scurvy killed more sailors than did warfare, accidents, and other causes. For instance, of Vasco da Gamma’s crew of 160 men who sailed with him around the Cape of Good Hope in 1497, 100 died of scurvy. In 1601, an English sea captain, James Lancaster, conducted an experiment to evaluate the effectiveness of lemon juice in preventing scurvy. Captain Lancaster commanded four ships that sailed from England on a voyage to India. He served three teaspoonfuls of lemon juice every day to the sailors in one of his four ships. These men stayed healthy. The other three ships constituted Lancaster’s ‘control group,’ and their sailors were not given any lemon juice. On the other three ships, by the halfway point in the journey, 110 of 278 sailors had died from scurvy. So many of these sailors got scurvy that Lancaster had to transfer men from his ‘treatment’ ship in order to staff the three other ships for the remainder of the voyage….
Not until 1747, about 150 years later, did James Lind, a British Navy physician who knew of Lancaster’s results, carry out another experiment…
not until 1795, forty-eight years later… was [scurvy] immediately wiped out…
Why were the authorities so slow to adopt the idea of citrus for scurvy prevention? Other, competing remedies for scurvy were also being proposed, and each such cure had its champions. For example, Captain Cook’s reports from his voyages in the Pacific did not provide support for curing scurvy with citrus fruits. Further, Dr. Lind was not a prominent figure in the field of naval medicine, so his experimental findings did not get much attention.

”

There was another unsolved problem Britain was facing during the 1800s in the wake of the Felling mine disaster: how to design a safety lamp that wouldn’t lead to explosions.  The most eminent British chemist at the time was Humprhy Davy at the Royal Society, who had discovered several elements and was a spectacularly popular lecturer.  As Richard Holmes relates in the Age of Wonder, the accidents committee approached Davy to help them understand the problem.  He began analyzing the gas (known as fire-damp at the time) and visiting mines and speaking with miners and overseers.  Upon returning to London, he “summoned Faraday to his assistance” and ordered an “apparatus capable of withstanding an explosion.”  He soon discovered that “explosions only occur[ed] when methane reacted critical ratio of gas to air (1:8 parts),” and eventually realized that a fine gauge iron mesh/gauze surrounding the flame would prevent explosions by providing a high surface area for cooling.  He refused to patent his safety lamp, and having “subdued this monster” and “scourge of humanity” to “much public gratitude,” thank you letters poured in from miners across the country; some of the letters were drafted by mine owners, but “the signatures were genuine… 47 [miners] were illiterate and simply put ‘x’ against their names.”

Problem solved, right?  Fast forward over 200 years.  A coal mine collapse earlier this year in Pakistan killed over 40 people when methane gas in the mine ignited and caused an explosion.  But what’s the root cause?  Methane doesn’t auto-ignite until over 500 degrees Celsius, and neither does coal dust below 425 Celsius.  That kind heat could have been generated through friction during mining activities, or even by but coal dust in a pile that oxidizes, leading to a runaway exothermic reaction.  But I think the picture below (taken recently) gives us a clue.  Some of the miners are still using kerosene lamps when they don’t have access to hard hats with electric lamps.  Why would anyone still be using kerosene lamps, when safe electric lamps have been available for almost a hundred years?  Because the people using the lamps are not the ones buying the lamps – the prototypical agent problem, as this article details:

“

a major reason why working conditions are bad at the mines lies in the ownership and operational structure of the mines. The mine owners do not actually own the mines; they are long leased from the government. They are then sublet to contractors, delegating the whole responsibility of operating the mine with all the risks involved…Since the contractors are not the legal operators of the mines, the law cannot hold them responsible for the absence of safety measures and equipment as well as accidents.

”

The kerosene lamps have a lower up-front cost than the battery-operated electric lights, so the mine owners/operators would favor buying them for the miners.  Even if the miners themselves know about electric lights, they’re simply too poor to afford them on their own.

A Look Inside a Busted Gas Turbine

I was touring a 300 megawatt natural gas combined-cycle power plant last year, when my friend mentioned that one of the turbines had been replaced due to damage. “Oh – what happened?” As it turns out, the engineer who was testing the turbine began spinning it up without checking the lubrication – which was non-existent. Once the RPMs got high enough, the ball bearings couldn’t take the friction and melted away, leaving the rotor to rattle around on its shaft and the blades to scrape against the casing. Insurance paid for the replacement, while the engineer managed to keep his job – and get schooled.





ouch:



The blades are made from single-crystal nickel superalloys with a thermal barrier coating. Rolls Royce has some interesting exploded diagrams and time-lapse construction videos of their jet engines (which are based on very similar technology) here. Single crystals have found other applications in clean energy – from silicon wafers for solar cells to massive, fast-growing KDP crystals for fusion at the National Ignition Facility.