Quarterly Newsletter #4
April to June 2025 | Notes from the Solar Industry
This quarter, I found myself chatting a lot about AI with an old classmate who was a bit of a genius during my undergraduate days of electrical engineering. The genius was not because he went on to earn his PhD or holds numerous patents, but because he was so deeply in love with the subject of Digital Signal Processing that even as a 19-year-old, he could explain very complicated concepts in simple words.
In all of our conversations, he brushed aside most of AI as noise and told me that its most fundamental level, it was basic mathematics done right, over a large dataset.
It then seems apt that, when I recently read a WSJ article titled “It’s Known as ‘The List’—and It’s a Secret File of AI Geniuses”, there was one line in the article that caught my attention and it said, “the people who get notes from Zuckerberg have a few things in common. They need to know calculus, linear algebra, and probability theory …’
AI in solar is deeply interesting because, at the most fundamental level, its image processing repeated over a large dataset. There are two use cases which I thought were very interesting.
Downstream Use Case - On-Field Drone Thermography
Drone thermography and visual capture are well-known field assessments. I spoke with a firm that does those tests. While the capture is done through drones, the failure analysis is done manually by engineers. They have recently raised money to build their AI engine and train it. The critical learning is that while a lot of the data is being created, it is fragmented and not consistent enough to analyze. A recent report from BCG called “A New AI Playbook for Renewables Energy Companies” supports this view.
Given the above, perhaps it’s important to build the right data and technology-first culture within renewable energy firms. A CEO of a leading solar manufacturing company in India told me that sector-specific software interventions are a big white space. He said he had to internally build a Manufacturing Excellence System (MES) as there was no readily available software in the market. With 75 GW of solar deployed in India and probably the world’s largest software talent pool in India, I would have thought there would be many options for him but seems like there is a wide gap (hence a huge opportunity) between product x industry sync. Since MES is an enterprise-wide intervention, I spoke to an experienced consultant who works for a global ERP firm. He said, “We have the products, but we are looking for use cases within the industry.”
Two more points before I go back to drones:
My experience suggests that culturally most solar/renewable companies still operate as old-school ‘power sector’ firms, and technology refers to TOPCon and HJT. Perhaps, it’s important to enable basic technology infrastructure to channel their existing and ever-increasing data into insights.
Unfortunately, the last quarter saw a few wars, but in today’s age, not all wars are physical. Renewable energy is a significant part of India’s energy mix now then it seems worrying to me that a recent industry journal showed the top five solar inverters from just one country. Inverters are intelligent devices that can be controlled remotely. Data security is more important than ever before.
Back to drones, that firm is now feeding the AI engine with lots of data, on what is right and what is not. The hope is that the model gets intelligent with time and the analysis requires less human intervention, thereby significantly decreasing turnaround time for the customers.
Going back to my friend’s analogy - two large matrices multiply with each other and eventually you will have 3-4 outcomes with coefficient i.e. weights. The more the model is trained, the stronger will be one of the weights, hopefully the right one. The learning from the drone thesis is that while AI is great, there are some conditions preceding its success. A large dataset is not enough; it has to be secure, usable, and repeatable.
Upstream Use Case: Inline inspection of solar modules to increase first-time quality
I quite enjoy visiting solar module manufacturing plants - takes me back to the early days of my career where I spent countless hours on the automotive manufacturing floor. During those days, EOLFTQ was a terrorizing number, but we will get to that acronym in a bit.
Nowadays, manufacturers talk of AI-powered solar module lines, which got me curious. I noticed that it’s less AI and more VI (visual inspection); a camera takes an electroluminescence image (think X-ray). Over time you take many pictures and make sense of them. A typical module will undergo four EL images across the assembly line:
After the stringer - while it’s still not a module
Pre-lamination (pre-lam)
Post lamination (post lam)
Final check
Some numbers:
A 1 GW module line running for 330 days/year produces 3 MW a day or 1500 modules per day (assuming 500 Wp/module).
Even if each module gets clicked twice i.e. pre and pre-lam EL image, that’s 3000 images a day or ~ 1 million images a year
… and that’s only 1 GW. India has 75 of them. Globally, 1800 by the end of 2025.
A notable observation in the solar circles is that factories in countries that are “new” to solar manufacturing take time to get their manufacturing process right. One observes more quality control issues from such countries. EOLFTQ stands for End of Line First Time Quality - a metric used to measure how much you got right on the first very time. Solar module manufacturing is a highly commoditized, high volume low margin business, therefore it’s imperative to get the FTQ right. If your FTQ is poor, your product will go through rework - the challenge with rework is that:
It is a waste of time and time is money.
The rework is often done manually by operators who are not well-trained resulting in real losses i.e. reduction in overall line yield. That’s not the operator’s fault per se - there is a learning curve in any industry and the attrition in the solar module last mile operator workforce, I am told, is very high!
I have heard lines having 5 - 40% reject rates in India. I am told the median is somewhere between 8 - 9%.
Even a 1% saving in the reject rate would mean:
3000 modules/day x 500 Wp/module x 1% benefit in reject rate x 20 USD cents / Wp = $3000 a day or $1m a year.
One could argue that not all rejected modules go unsold but even then, small changes can bring meaningful impact on the overall line yield as the volumes are so large.
To verify my understanding, I called two manufacturers who told me that all the AI is already available but often the operators manually shut the AI intervention because the reject rate was too high. There could be two plausible explanations:
The reject rate is high because the modules are genuinely poor but one doesn’t want to stop the line - manufacturers are very busy, so don’t stop a running line! This will create problems in the future.
OR
This AI model is showing too many false positives - this goes back to the need for training the model with the right outcomes.
In conclusion, there is tremendous potential for AI in solar but two big learnings:
Any repeatable task can potentially utilize AI - to optimize the process, to reduce costs - however, we have to go back to the fundamentals of getting the right quality of data and training the models the right way and that takes time.
Understanding the specific needs of the industry is critical - Speak to the customers before building your software!
Speak more? If you find this newsletter interesting and want to speak more about AI within the solar industry, please message me on LinkedIn.
Disclaimer: Views expressed are personal and not those of any organization that I am associated with.