Cost Modeling for Specific Photovoltaic Technologies (Text Version)
This is the text version for a video—Cost Modeling for Specific Photovoltaic (PV) Technologies—showing reference designs, process flows, and cost models NREL analysts use for PV technologies.
It’s Part 2 of NREL’s Solar Techno-Economic Analysis (TEA) Tutorials video series.
Now, I'll walk you through some cross models we have for crystalline silicon photovoltaics. I'll use this section also to highlight typical assumptions for cross models and reiterate some of the points already covered in the methods section.
Crystalline Silicon PV Supply Chain
Here I'm showing the primary supply chain for silicon PV. I'm going to use this slide to briefly discuss scoping. You could model any or all of these stages in different degrees of detail or you could go even further up the silicon supply chain and model production of ferrosilicon and metallurgical rated silicon. For our model, we limit it to these five stages though we do include multiple scenarios for polysilicon and then both mono and multi crystalline for the other four stages.
The primary goal of our TEA here is to benchmark costs for the industry and Department of Energy so it can aid business and policy planning and also give some insight into cost categories that can help direct research. I'm showing the typical plan size for each of these products here because we want our model to reflect these output scales so that it's representative of what's being done in the market currently.
Costs by Step for Siemens Polysilicon Production
Starting with polysilicon, our graphics team made nice animations of the manufacturing process for all of the other steps. Unfortunately, I don't have one for polysilicon. So, I'll just jump right into some of the results. These results are a little dated, but we have more recent results for the rest of the steps so just keep that in mind.
Here I'm showing costs by step for polysilicon produced in the United States using the Siemens process, which is the most common method of polysilicon production. You can see the first step is producing trichlorosilane from metallurgical grade silicon or MG silicon. As I just said, we scoped our TEA so that we aren't doing a separate model for MG silicon. We just use an input price for MG silicon. In these results here, it's clear that some steps in the production process are more expensive than others where some have higher material costs. Others have higher electricity costs or equipment costs.
Polysilicon Production Scenarios
So, then, the cost by step results are summed into total costs and price estimates for a range of scenarios. The cost by step data for the Siemens process shown on the last slide is summed into the first bar on this chart. The Siemens steps were modeled using data from multiple China locations as well as the United States data. You can see the differences between the first three bars here occur mainly in the electricity and labor categories. Two other polysilicon production methods were modeled for these different locations as well, including the hyperpure Siemens process as well as the fluidized bed reactor process. Once I move onto ingot production in the next slide, it will be clear why it's important to both model the Siemens and FBR processes.
Process Flow for Monocrystalline Silicon Wafer Product
Now I'll just review the manufacturing process for monocrystalline silicon ingots and wafers. The silicon feedstock is typically polysilicon chunks manufactured via the Siemens process and small polysilicon granules produced by the fluidized bed reactor process to fill in the gaps between the Siemens chunks. These polysilicon products are loaded into a cylindrical crucible which is then heated so that the polysilicon melts and is then doped. A monocrystalline silicon seed crystal is introduced and then rotated while pulling upward to begin the crystal pulling process. Pulling continues until most of the polysilicon has crystalized into a single crystal ingot. Some polysilicon is left in the bottom of the crucible as it now has a higher concentration of impurities.
The ingot is removed from the crucible and segments are cropped off either end of the cylinder. The rounded exterior of the cylindrical ingot is cropped in sections to create a pseudo square ingot. All the cropped sections are etched and readded to the crucible for use in the next ingot. The sharp corners of the pseudo square ingot left from sawing then undergo grinding and polishing. The pseudo square ingot is then glued to a glass substrate on one face. A multiwire diamond wire saw has been used to cut the entire ingot into wafers simultaneously. The wafers are still glued to the glass substrate after cutting so the glass is placed in a chemical bath to dissolve the glue and release the wafers. The wafers are then cleaned, separated, and inspected.
Breakdown of Costs by Step for Monocrystalline Wafer Production
Here are the costs by step results for a monocrystalline P-type wafer production. You can see a lot of the information about the reference design and scenario details are listed in the title and other input data is shown throughout the figure. From these results, it's obvious that ingot growth is more expensive than wafering and it's very energy intensive.
I also want to use this slide to start talking about the concept of yield. You'll notice here that the wire sawing step results in five percent of wafers breaking. This means if you need 100 wafers, you're actually going to have to set up production for 105 wafers. This will compound in later steps. I'll come back to this number.
Monocrystalline Ingot and Wafer Production
Here are the total costs for monocrystalline wafer production and the estimated minimum sustainable price. Our wafer cost model was recently updated to model production of M4 format wafers which are larger than the M2 format wafers that were previously used for the past five years or so, but the industry is now transitioning to M4. There are economies of scale across many cost categories when producing, processing, and assembling modules from larger wafers.
Illustration-Only Device Stacks Under Development
Now that we have wafer production modeled, we can then proceed to cell fabrication or cell conversion. Here I'm showing our reference designs for five cell types. I'm not going to discuss all five in the interest of time but I just want to note that these architectures vary in terms of their wafer polarity, efficiency, current market share, or scale production, and other features, such as bifaciality, which offset their TEA performance. The PERC cell or passivated emitter and rear cell is currently the most widely produced cell type. So, I'm going to briefly [Audio skips] the PERC conversion process in the next slide.
Process for PERC Cell Conversion
The first step in the PERC cell conversion process is to scan the P-type silicon wafer. Then, any saw damage is removed. The surface is texturized. It goes through a pre-diffusion clean. Then, diffusion of N-type phosphorous dopins is completed using POCI gas. Lasers are used to create a selective emitter. Then, a wet chemical etch is used to remove the layer of phosphosilicate glass that formed during POCI diffusion. The rear side is then planarized and then isolated by etching phosphorous from the rear.
Next, high temperatures are used to form a layer of silicon dioxide. Then, aluminum oxide is deposited by plasma enhanced chemical vapor deposition or atomic layer deposition for surface passivation. Next, silicon nitride is deposited by plasma enhanced chemical vapor deposition which serves as a front side anti-reflective coating, a backside reflector, and overall surface passivation. Lasers are then used to create openings in those oxide and nitride dielectric layers so that only contacts can be created between silicon and the rear metal, which is about to be added.
The next step is then to screen print silver paste for tabbing on the rear side and aluminum paste for a back-surface field on the rear side. Silver paste is also used to screen print the front side for fingers and possibly busbars. All the screen-printed pastes are then cofired. Finally, there is an optional hydrogenation step that can improve efficiency and passivation or cells can proceed directly to current voltage measurements, visual inspection, and classifying cells into bins typically based on their efficiency or visual features.
Costs by Step for PERC Cell Conversion
Here are the costs by step for the PERC conversion process. Again, we're showing a lot of the input data on this figure. You can see we've broken out metallization material costs in gray since they represent such a large fraction of total costs, which is pretty typical for silicon conversion processes.
Here I want to pivot back to discussing yield again. When the wafers are scanned at the beginning of this process, around .1 percent are rejected for quality purposes. There is a similar rejection rate through each step of the rest of the process. By the end of the process, about one percent of wafers or cells are ultimately rejected. Again, to go back to our 100 number from the wafer yield discussion, if you need 100 cells, you actually need to aim to produce 101 cells. So, when before, I said you'll need to produce 105 wafers to get 100 wafers, you now need to produce 107 wafers to get 100 cells. You can apply this to any other material in the process and further up the supply chain to polysilicon production as well, but I just wanted to use wafers as a discrete example.
Now, just to provide an N-type cell example before I move into total cell cost results, I'll review the process flow for our silicon heterojunction solar cell conversion, which is the next cell type expected to gain the most market share in the near future.
Process Flow for SHJ Cell Conversion
The first step in the silicon heterojunction cell conversion process is to test the incoming N-type wafer. Next, any saw damage is removed, and the surface is texturized. Then the wafer is dipped in hydrofluoric acid to remove any oxide. Then, hydrogenated amorphous silicon layers are deposited on the front side by plasma enhanced chemical vapor deposition where an intrinsic layer is deposited first, followed by a P-type layer.
The same is completed on the rear side where hydrogenated amorphous silicon is deposited, except an N-type layer is deposited after the intrinsic layer. Then, indium tin oxide is sputtered onto both the front and back sides and the front and back sides are then metallized where, notably, the back side has periodic contact such that the absorber is exposed. Finally, the completed cells undergo current voltage testing and sorted into bins.
Costs by Step for SHJ Cell Conversion
Here are the costs by step results for the heterojunction or SHJ cell conversion process. Again, the metallization materials are a large fraction of the costs and for SHJ production that equips any other production step. I do want to note that our cost by step analysis shows only costs that are specific to the cell conversion process. So, the wafer price is not shown in these figures. I do want to highlight wafer pricing in the next slide.
Comparison of Cell Conversion Costs
Here are the total costs and MSP estimates for a few different reference designs and scenarios. You can see here we're comparing multiple iterations of PERC and SHJ architectures. The PERC rear contact can be printed to create either monofacial or bifacial cells where less metal is used on the back surface for the bifacial configuration and some cost savings occur due to decreased metal usage. For SHJ scenarios, when comparing screen printing versus electroplating metallization processes, the difference in metal usage versus equipment costs essentially cancel each other out. When comparing screen printing of busbarless versus busbar contact configurations, the reduction of metal usage in the busbarless configuration is combined with the lower equipment costs of screen printing and results in the lowest cost SHJ cell scenario. It is also important to note here that the PERC cells use P-type wafers, which are less expensive than the N-type wafers used for the PERL and SHJ cells shown here. However, the wafer costs for the SHJ cells do not appear as elevated as the PERL cell due to the higher SHJ efficiency and using the metric of dollars per watt.
Now we'll move to the fast age of our cost model module assembly. I'll review a typical monofacial PERC module assembly process in the next slide.
Process Flow for PERC Module Assembly
The first step is the automated unloading of AR coated solar glass where the AR side is facing down. An encapsulant sheet is then laid down on top of that glass. Strings of cells are created by automated tabbing, stringing, and contactless infrared soldering where the silver on the front side of the cell is connected to the silver on the back side of the adjacent cell. Robotics then place these strings onto the encapsulant sheet and string connector ribbons are soldered. After this, a second encapsulant sheet is laid down on top of the strings and a backsheet is then laid on top of the second encapsulant sheet.
The module then proceeds through visual inspection and electroluminescence testing. Once it passes quality control, the module is then laminated in a furnace. After lamination, the edges of the module are trimmed, and an aluminum frame is fitted, and silicon sealant is applied. Next, string connector leads are soldered into a junction box. A serial number is affixed to the module and the module undergoes curing. Finally, the module undergoes current voltage testing and electroluminescence inspection after which the modules are sorted into bins based on performance.
Costs by Step for PERC Module Assembly
Here are the costs by step for a monofacial module assembly. We've broken out most of the individual material costs for these steps where the two largest cost contributors are the front glass and the aluminum frame. It's important to note that the glass is more expensive than the back sheet shown here, which will be relevant when comparing monofacial versus bifacial scenarios in the next slide.
Example Cost Model Results for Different PV Technologies
Here we're showing total costs and MSPs for the appropriate module configurations that correspond to the cell scenarios we discussed a few slides back. Note that the module price trends do not exactly track with the cell price trends from the previous figure. In this figure, bifacial PERC modules are more expensive than the monofacial PERC modules due to the rear sheet of glass replacing the less expensive back sheet.
Road Map for Monocrystalline Silicon PV
Finally, here is an example of how you can use your cost model and TEA results to roadmap the technology. The left most bar here is showing the TEA results generated a few years ago to benchmark the industry at the time. The next bar to the right shows our cost model results from last year. The data updates that resulted in the cost changes between these two years are summarized above in the text box. These are mainly material cost reductions, material usage reductions, economies of scale, and efficiency improvements. We then estimate what we think the next data updates will be based on the trends we are observing in industry to create a projection for the coming years, including larger wafer and ingot sizes, further material usage reductions, as well as efficiency and throughput increases.
That concludes my silicon section review. If you're interested in seeing more detail on any of this, please see our most recent publication where our silicon model results were published in great detail. The reference is shown here on this slide.
Thin-Film Cost Modeling Results
Now we will go through some of the thin-film cost modeling results for CIGS, CdTe, and III-Vs.
CIGS Device Stacks
We're going to start with CIGS. As you know from the earlier presenter, the first step in technomagnetic analysis is to establish a reference device stack that specifies the layers and the thicknesses that are used in the device. For CIGS, unlike, for example, a specific class of silicon technology, the device stacks are a lot less standardized and there are many different designs. I'm showing you three on the bottom that we've used for benchmarking in the past. As an example, the one that I'll be walking you through today in terms of cost is the one on the far right with the cad sulfide buffer layer.
When we're creating device stacks or if you want to do this on your own, there are really a few key things that are important to consider. You really want to make these so that you're benchmarking the technology and its costs without being specific to one company and so that you can generally inform R&D directions. There are a few reasons why it's really important not to try to represent one specific company. This helps protect proprietary information. It's really important that we maintain trust with the manufacturers who we collect data from and ensure that their proprietary information is kept confidential so that we can continue to have those relationships and collect the data that we need for our analysis from them. Having a more generic benchmark design also helps avoid the appearance that you are representing a specific company's cost or are trying to market that company. It also just makes the analysis more broadly applicable with an industry in academia.
Once we've established the device stack, the next step is to establish a reference manufacturing process flow again as you also saw with the other technologies. So, now, I'll walk you through that for the CIGS process we're analyzing here.
CIGS Reference Manufacturing Process Flows
Now I'll walk you through the manufacturing process flow that we use to calculate the CIGS costs in this case. First, unload, inspect, and wash the back glass, which is soda lime back glass in this case. Then sputter the molybdenum back contact layer. Then laser scribe through that molybdenum back contact layer. Next, deposit the CIGS layer itself. In this case, we're assuming a two-step sputtering plus batch selenization and sulfurization process. Then deposit the buffer layer either with sputtering or chemical bath deposition. Then scribe again using a mechanical scribe through the buffer down to the molybdenum. Then deposit the window in TCO layers using sputtering in this case but MOCVD or other processes could also be used. Then scribe again down to the molybdenum and use laser edge isolation. Finally, electrically connect the monolithically integrated cells by solder welding metal ribbon busbars and connecting adhesive tape. Then assemble the module by connecting the busbar assembly to the front glass with the EVA and an edge seal. Lastly, attach the junction box, light soak, and perform any testing, including hi-pot testing, ground continuity, and solar simulator J-V testing. Then, visually inspect module, bin it, and pack it.
Other Key Input Assumptions
Finally, I'll walk you through the other key input assumptions we made in our model. All these inputs are really important in calculating costs. If you're setting up your own model, it's important to carefully select each of these values and think about why you're selecting them. We do this based on market benchmarks, so, where we think that the leading companies are in terms of their capacity yield loss and where they're manufacturing. We also change these parameters to explore different scenarios in terms of future directions for technologies or to look at variability between companies.
Here we assumed a one gigawatt per year production capacity for the plant at a ten percent cumulative manufacturing yield loss. In our models, we actually have a different yield loss for each step that's based on our interview data. And then, it's cumulated to determine how much is lost at the end of the production process. We're assuming manufacturing in Japan, which is where Solar Frontier, who is the leading CIGS company in terms of their capacity, is located. Again, we're not trying to represent Solar Frontier's cost with this. It just helps us get a benchmark for where CIGS may be generally with that location of manufacturing.
CIGS Step-by-Step Costs: 2019 Results
Here's the result. These are the step-by-step costs for CIGS. This is based on 2019 data. We haven't updated these results this year. You can see a couple things that jump out from this. One of them is that you add up all of the costs associated with the balance of modules, so, things like unloading and washing the back glasses. Materials costs includes the cost for the glass itself.
The front glass, bonding everything together, the junction box, and sticker label, and the frame actually add up to being a significant portion of the overall costs and aren't really related to the cell layers themselves. Another thing that you'll notice is this selenization and sulfurization step is very expensive. This is because of the batch process that we assumed, which is very slow. Because of that, there are many tools that are needed in parallel to meet the required throughput. That increases the depreciation expense as well as the utilities and labor costs.
CIGS Cost Reduction Road Map: 2019 Results
If you add up all of the bars from the step-by-step chart on the previous slide, you'll get around 40 cents a watt. That's just for the manufacturing cost. That does not include any overhead costs or markups. We wanted to look and just see on a manufacturing cost basis where could this technology potentially go in the future if there are different improvements made. That's what this waterfall chart is showing you starting at that 40 cents a watt value. What happens if you increase the manufacturing yield from 90 to 95 percent? You can see that has a little bit of an effect here.
A much larger effect could come from increasing the throughput of manufacturing the CIGS layer. In this case, that's the combined throughput of sputtering the CIGS precursors as well as the selenization and sulfurization step. But, the vast majority of the time is spent in that batch selenization and sulfurization process, like I mentioned. So, it's really about speeding up that process or finding an alternative. There are alternatives, such as co-evaporation or inline rapid thermal processing for selenization and sulfurization that people have looked at. We have seen that those can help to improve the throughput of these processes but are not at the scale of the two-step batch process yet in production.
If you remove the frame, you can also save a pretty significant amount in costs. There are some frameless modules available today as well. You can see how increasing the efficiency in different steps influences costs. Efficiency can really be a big driver of cost reductions as well.
Another major way that we've seen you can decrease CIGS cost is by increasing the area of the module. In this case, the example is shown for a 2X increase in module area. If you add all of these together, you could end up at some module with a manufacturing cost of around 16 cents a watt DC. Again, that's not the price you would see. It does not include overhead, and it doesn't include any profit. That's what you're looking at here.
Representative Device Stack for CdTe
Now we'll take a look at something similar for CdTe. This is the representative device stack that we use for CdTe cost modeling. We assume typical cell widths between 5.5 and 5.7 millimeters. The CdTe layer itself is between two and three µm.
Process Flow Used for NREL’s CdTe Cost Model
Now I'll walk you through the process flow that we use for NREL's CdTe cost model. First, we deposit a transparent conducting oxide or TCO on heat treated front glass. Then we wash the front glass. Then deposit a cadmium sulfide buffer layer followed by the CdTe absorber layer and then a cad fluoride activation. Then we have a precontact wash and aqueous copper diffusion step followed by a P1 laser scribe through the cad sulfide, CdTe, and TCO layers. Then fill those scribes in with insulator for cell isolation. Then we have a second laser scribe. Then screen print back contact metallization paste. Then sputter an adhesion layer and a diffusion barrier. Then sputter the back-electrode stack.
Then have the third scribe. This is typically a mechanical scribe down to the CdTe followed by edge isolation. Then we electrically connect the monolithically integrated cells by solder welding metal ribbon busbars and conducting adhesive tape. Finally, assemble the module by feeding the busbar ribbons through the predrilled hole in the tempered back glass bonding the cells and the busbar assembly to the EVA back glass and edge seal. We complete the assembly by attaching the junction box onto the module, light soaking, and then performing any testing, hi-pot testing, ground continuity, and solar simulator J-V testing and then visually inspecting the module, binning and packing it.
III-V Solar Cell Cost Modeling
Now we'll take a look at III-V solar cell cost modeling. There are many different III-V solar cell device stacks as well, including variations on cells with different numbers of junctions, as well as just having many different cells with single-junction technology, dual junction, or three plus junctions. What I'm showing here is the reference device stacks that we use for a three-junction device, which is a lattice matched device on a germanium substrate. Again, there are many different junction devices and architectures that are used today but this one is one of most common. The dual junction device in the middle, which is a gallium arsenide cell with gallium indium phosphide on top and then a single junction gallium arsenide cell.
In our benchmark case, we assume that all these technologies have their epitaxial layers grown via MOCVD, which is the most common process used in the industry today. We also assume low volumes consistent with where the industry is today. Currently, on the order of hundreds of kilowatts of III-V solar cells are produced each year, which is many orders of magnitude less than CIGS, CdTe, and particularly silicon. The main area where these are used is in space applications, which are less sensitive to the cost per watt, which is high for these technologies but really need the extra performance in terms of efficiency and weight in order to optimize the overall cost of the space structures themselves.
1J and 2J Cell Costs with 5 Substrate Reuses
First, we'll look at our results. We'll look at the costs associated with those single and dual junction device stacks that I showed earlier. In this case, we're assuming that the cells are epitaxially lifted off of the gallium arsenide substrate and then that the substrate is reused five times. This is different from the assumption that we made in the three-junction lattice matched on germanium cell we just looked at. That's the primary reason for the difference in the substrate cost between those two device stacks that I'm showing.
Currently, the yield associated with that epitaxial lift off and reuse process is highly uncertain and variable. So, that's something to just take with a grain of salt. We made an assumption here. You can see the details of the assumptions that were made in our most recent publication on this topic if you're interested. What you can see from these results is that even with the substrate reuse, the costs are still pretty high. Manufacturing costs around $40.00 a watt. MSP is between $60.00 and $70.00 a watt for both technologies. The higher efficiency of the two junction offsets is the increase in the epitaxial growth costs associated with that because the efficiency also helps to drive down the non-epitaxial growth costs of the cell. In both of these cases, you can see though that substrate processing and epi growth all still play an important role in driving the costs.
The costs associated with CMP or a chemo mechanical polish that's used during the reuse process to reprepare the surface of the substrate for the next growth is a very significant cost and really limits the degree to which you can drive down substrate costs unless you're able to find an alternative to CMP. Again, like any of the other technologies, costs for specific company and prices that that company is offering in the market, which can be separate from costs completely, may differ.
Challenges of Estimating III-V Costs at Scale
Now, I want to talk to you a little bit about some of the challenges associated with trying to take these estimates that we have of costs at these very low production volumes and translate those to a high volume to understand the potential of these III-V technologies. Like we mentioned before, several hundred kilowatts a year is the current volume that these cells are produced at. So, going to a hundred megawatts a year would be greater than 100X increase in production volume. A gigawatt per year would be greater than 1,000X increase in production volume. So, there are a lot of assumptions and guesses that we have to make in understanding how these costs could scale.
Just to give you one example of that, the current global production for gallium arsenide wafers on a six-inch equivalent basis, if you assume a 30 percent efficiency for a cell, is around six megawatts per year. That's for all applications, not just solar. Solar is actually a relatively small part of it. But this is including all applications of gallium arsenide. That means that the whole supply chain would need to expand significantly to serve these markets if solar cell production scaled up to these levels. Because of that, the cost and price are uncertain. If we ask the manufacturer, what is this going to cost at that volume? They're guessing, too, because they haven't really developed the processes to do that. They're also not sure how that would influence the dynamics of the input material pricing.
We try to do as good of a job of this as we can, again, by really interviewing manufacturers and trying to dig into what drives differences in input material pricing. What are the commodity prices of the input materials? How much of those are produced each year and are needed to create the different components that go into the cell manufacturing, like the substrate.
Estimated III-V Costs at Scale with Future R&D Progress Implemented at Scale
If we do try to guesstimate what the cost could look like scaled to the best of our knowledge using a combination of our model, and those interviews, and fundamental data about pricing, and how that evolves with scale, this is what we could get. Again, in our most recent publication, there's a road map between those very high $50.00 a watt plus costs and this stage one bar. If we jump ahead to a case where we're looking at 50 megawatt per year production – which is still very modest in terms of solar production generally – and we assume that instead of MOCVD we use dynamic hydride vapour phase epitaxy for the epitaxial growth, 25 substrate reuses with a lower cost for each chemo mechanical polish and that you don't need to use the CMP every reuse, only every five reuses, and that you can electroplate lower cost metallization as opposed to using higher cost metals and EBM process like today, you could get around $2.00 a watt. Again, that assumes that a laboratory process dynamic HVPE is scaled up successfully to higher volumes without unanticipated costs arising. If you get up to the gigawatt plus production volumes and you can get a six-inch or an eight-inch substrate that's $1.00 total for everything – that's the substrate, any cost associated with lift off if you're still doing that, any cost associated with the chemo mechanical polish if you're still doing that – then you could potentially get up into the sub 50 cent a watt or 50 cent to a dollar a watt range.
That gives you an overview of the current technology for these three different thin-film materials, CIGS, CdTe, and III-Vs, and where they may go in the future.
To continue with Part 3 of the Solar TEA Tutorials video series, see PV and Storage System Cost Benchmarking (Text Version).