The Purdue team that proved you can predict the planet’s solar energy yield from just five well-chosen locations — using physics-guided machine learning.
Ask any solar developer the first question of a project: how much electricity will this site actually produce? The answer, traditionally, comes from physics-based simulation — PVSyst, PVMAPS, and their siblings — run by expert consultants on proprietary software, over weeks of work.
It works. But it's slow, expensive, and — here's the catch — it depends on data from a handful of highly-instrumented test sites in a few rich countries. There's no global experimental network for solar. A manufacturer launching a new module can't afford to install it in a hundred locations for a year of testing. And commercial operators guard their performance data like trade secrets.
A team at Purdue University (Jabir Bin Jahangir and Muhammad Ashraful Alam) just demonstrated something that quietly rewrites the rules of this game: you can predict the annual energy yield of solar farms anywhere on Earth — within single-digit percentage error — using data from as few as five locations.
Five. On a planet of 510 million square kilometers.
The Core Insight: Physics Picks the Variables, Climate Picks the Sites
The Purdue approach — physics-guided machine learning (PGML) — starts with a deceptively simple question: what actually determines a solar farm's energy yield?
Their answer: three meteorological variables.
- Global Horizontal Irradiance (how much sun hits the ground)
- Ambient temperature (panels lose efficiency when hot)
- Clearness index (how cloudy it is — the split between direct and diffuse sunlight)
That's it. Everything else — module temperature, latitude, longitude — is either correlated with those three or redundant. A neural network mapping these three inputs to monthly energy yield captures the physics, because the physics was used to choose the inputs.
But three inputs isn't the clever part. The clever part is where you collect the training data.
PVZones: The Climate Map Nobody Made Before
If you want to train a model that works everywhere, you need training data that represents everywhere. Randomly sampling locations is inefficient — 90% of your samples land in climates you already understand.
So the Purdue team built PVZones: a climate map of the world, created specifically for solar energy yield. They took NASA's global climatological data, extracted their three physically-relevant variables, and ran k-means clustering.
Result: the world divides into seven PV-specific climate zones — latitudinal bands, symmetric about the equator (naturally — the sun doesn't care about borders). The Sahara and the Australian outback share a zone. So do the UK and southern Chile. The US and China, being enormous, each span several zones — meaning one country can sometimes host the whole world's climates.
This is the key that unlocks everything: a site in one zone represents millions of square kilometers elsewhere. Data collected anywhere is data collected everywhere — as long as it's representative, not random.
The Five-Location Miracle
Here's where the results get genuinely astonishing.
Experiment 1: Five sites, five different zones. The team simulated monthly yield data from just five locations, one in each of five PV zones (60 total data points), then trained the model. The model then predicted energy yield for the entire globe at 0.5° × 0.5° resolution — that's ~180,000 grid cells.
- 95% of the world's locations predicted within 9 kWh/m² of the physics benchmark.
- Overall root mean square error: 3.85 kWh/m² — roughly 1.5% of a typical annual yield.
Experiment 2: Five US test sites. They repeated it using only the locations of DOE's regional PV test centers (Sandia, NREL, and friends) — all in the continental US, none in the tropics or southern hemisphere. Result: still 95% of the world within 16 kWh/m², RMSE 7.62 kWh/m². Errors concentrated exactly where you'd expect — in the zones with no representative in the training set.
The takeaway is almost philosophical: a small, diverse, physically-relevant dataset beats a large, random one. Every time.
The Messy Reality: 3,000 Crowd-Sourced Systems
Of course, controlled experiments are one thing; the real world is another. So the team did the hard version: they scraped over 3,000 real PV systems from public repositories (PVOutput, DuraMAT, university dashboards, literature) — systems ranging from a few kilowatts to megawatts, with unknown tilts, unknown orientations, and noise everywhere.
The heterogeneity was brutal. Raw yields spanned four orders of magnitude — a 5 kW rooftop and a 50 MW plant produce wildly different numbers. But physics came to the rescue again: they scaled each field system against a simulated reference system at the same location, homogenizing the data, and trained on it (36,000 monthly observations).
- Error below 10% for 95% of global locations — from messy, public, heterogeneous data.
Then they found the data's blind spot: the Arctic and Himalayan regions (PVZone 2) and other high-latitude zones had almost no installations — nobody builds solar farms where there's no sun. The model struggled there (15% error at the 95th percentile).
The fix is the paper's most elegant move — physics-guided data fusion: they injected simulated data from just 5 synthetic sites in the missing zone into the training set.
- Error drops below 5% for 95% of the world.
Finally, they validated against genuinely unseen field data: a site in Ohio (predicted within 0.9% of actual yield) and one in Cochin, India (predicted within 0.3%).
Why This Matters for Solar Due Diligence
For anyone evaluating solar projects — site selection, M&A, technology qualification — this is a toolkit-level change:
- Site screening without consultants. Rough yield estimates for any parcel on Earth, in seconds, from public data. That transforms early-stage screening: you can rank a thousand candidate sites before spending a dollar on detailed engineering.
- Technology qualification gets global. A module vendor can validate a new product's yield potential worldwide with a handful of test sites — strategically placed, one per PVZone — instead of an impossible global deployment.
- It's technology-agnostic by design. The framework doesn't care if the farm is fixed-tilt, single-axis, bifacial, floating, or agrivoltaic. New farm configurations and new materials (perovskite, tandem) just need their own PVZone-style calibration — the machinery carries over.
- Public data stops being a liability. The paper shows how to homogenize noisy crowd-sourced performance data into training-grade material — a recipe that works for any metric that depends on local climate (the authors note it extends beyond PV, e.g., reliability of automotive electronics).
The honest limitations: this predicts potential yield — degradation, soiling, and inverter losses are deliberately excluded; and the current model is monofacial-specific. But as a potential-yield atlas of the planet, it's unprecedented.
The Bottom Line
Purdue's paper contains one idea that will stick with you: the world is smaller than it looks. Because climates repeat across continents, a handful of strategically-chosen sites can represent all of them. Five locations predicted the planet within ~1.5% error. Three thousand messy public systems, homogenized and fused with a pinch of simulation, got under 5% — globally.
The days of "we can't know the yield without a month of PVSyst consulting" are numbered. Soon the answer will be: tell me the coordinates, I'll tell you the yield.









