How Climate Projections Work

A global climate model - climate scientists often call the core of it a general circulation model, or GCM - is not a statistical guess or a trend line drawn through past data. It is a representation of the physics of the climate system, translated into equations a computer can solve. The planet is divided into a three-dimensional grid of cells - imagine a mesh wrapped around the globe and stacked in layers through the atmosphere and down into the ocean. Within each cell the model tracks quantities like temperature, pressure, humidity and wind, and it applies the known laws of physics: how sunlight arrives and heat radiates away, how air and moisture move, how oceans store and carry heat, how ice reflects light, how clouds form. Crucially, each cell exchanges energy and matter with its neighbours, so heat and moisture flow across the whole grid the way they flow across the real planet.

The model then does something conceptually simple and computationally enormous: it starts from a known state and steps the entire grid forward in time, over and over, recomputing the physics at every step for years, decades, a century. Out of that vast calculation emerges a simulated climate - patterns of temperature, rainfall, wind and their seasons - that, when the physics is right, reproduces the broad behaviour of the real world. The key input for the future is the level of greenhouse gases, because that is what changes how much heat the atmosphere traps. Change that input and the whole simulated climate responds.

This is why a climate model is trustworthy in a way a mere extrapolation is not: it is grounded in physics we understand well, and it is tested by asking it to reproduce the climate we have already observed. It is also why it is not omniscient. Clouds, and processes smaller than a grid cell, cannot be resolved directly and must be approximated - and different modelling groups make those approximations differently, which is one reason models disagree. A GCM is a genuine, physically grounded window on the future, and an imperfect one - both things are true, and an honest designer holds them together.

The single most important word in this lesson is projection. A weather forecast is a prediction: it says what the actual weather will be at a place and time, and it is only skilful a week or two ahead, because the atmosphere is chaotic - tiny uncertainties in today's state grow until, beyond a couple of weeks, the specific weather is genuinely unknowable. A climate projection does not attempt that. It does not tell you the weather on a given future day; it tells you the expected statistics of the climate - typical temperatures, how often heatwaves occur, how the seasons shift - over a future period like the 2050s, and it does so conditionally: only if a particular assumption about emissions holds.

That conditionality is the heart of it. A projection is always an 'if-then': *if* humanity emits along this path, *then* the climate is likely to respond roughly like this. Because no one can know in advance how much the world will emit - it depends on politics, economics and technology, not physics - modellers cannot produce a single unconditional prediction of the future climate. They can only run the model under different assumed futures and report what each implies. So a projection carries its scenario with it, always; strip the scenario away and the number is meaningless.

There is a second reason projections come as ranges rather than points. Even under one fixed emissions path, running slightly different starting conditions, or different models, gives somewhat different results - the climate has natural year-to-year variability, and models approximate the unresolved physics differently. So scientists deliberately run ensembles: many runs, many models, whose spread describes how confident we can be. The honest output is therefore never 'the 2050s will be X'; it is 'across plausible runs and models, under this scenario, the 2050s are likely to be in this range'. Treating a projection as if it were a forecast - a firm number for a firm year - is the classic false-precision error, and it starts right here, at the difference between predicting the weather and projecting the climate.

There is a practical problem between a global model and a building. To simulate the whole planet for a century, a GCM must use grid cells that are large - often around a hundred kilometres across, sometimes more. That resolution is fine for the planetary picture, but a hundred-kilometre cell is far too coarse for design: it can blur a coastline, a mountain range and a city into a single average, and a building sits in the fine-grained local climate that such a cell cannot see. The gap between the model's coarse output and the site-specific information a designer needs is bridged by downscaling.

Downscaling comes in two broad flavours. Dynamical downscaling runs a higher-resolution regional climate model over a limited area - say the Indian subcontinent - taking its boundary conditions from the global model but resolving the terrain, coastlines and weather systems in much finer detail; it is physically rich but computationally expensive. Statistical downscaling instead uses observed relationships between the large-scale climate and local weather to translate the coarse projection into local terms; it is far cheaper but leans on the assumption that those past relationships hold in the future. Both aim at the same thing: turning a smeared global signal into something meaningful for a place.

Downscaling is genuinely useful - without it, global projections could not be brought to a building at all - but it is not free. It cannot conjure information the global model never contained; it can only redistribute and refine the signal, and it adds its own assumptions and errors on top. Dynamical downscaling inherits the errors of both the global model that drives it and the regional model that refines it; statistical downscaling leans on the assumption that past relationships between large-scale and local weather still hold in a changed climate, which may not be true. For India this matters especially: the monsoon, complex terrain, coasts and intense local heat are exactly the features a coarse cell hides and downscaling tries to recover, and exactly where methods differ most, so different downscaling choices can give visibly different local futures. So the local, site-level climate that finally reaches a designer has passed through another lens that sharpens the picture and, at the same time, adds another layer of uncertainty to everything the model already carried - which is why the file handed over should always be read as one plausible refinement, not the definitive local climate.

Put the pieces in order and you see a chain, and the honest point of this lesson is that uncertainty enters at every link. First, the projection depends on an assumed emissions scenario - a human choice no model can predict - and different scenarios diverge enormously by late century. Second, the global model itself is imperfect and approximates processes it cannot resolve, so different models give different answers even under the same scenario; the spread between them is real uncertainty, not a rounding error. Third, downscaling to a region adds a further method-dependent layer. And later, as the next lessons show, turning this into a usable weather file by morphing adds yet another. None of these links is a mistake to be fixed; they are irreducible features of trying to know an unknowable future.

This is why the output must be read as a range, not a number. Each link multiplies the possibilities: a handful of scenarios times several models times downscaling choices gives not one future but a fan of plausible futures, and near-term that fan is narrow while by late century it opens dramatically. A designer who takes a single downscaled file and treats its hourly values as *the* climate of 2050 has quietly discarded all of that - collapsing a wide, honestly uncertain range into one falsely precise line, and usually the mild-looking middle of it. The discipline the rest of this module builds is the opposite: to hold the range open, understand its direction and severity, and design for robustness across it.

And a boundary belongs here, at the start. Understanding how projections work lets you use them wisely and question them intelligently; it does not qualify you to certify a building's performance against them. The binding building-physics, energy, thermal-comfort and climate-risk engineering - and any compliance or life-safety determination - stays with qualified specialists using validated tools, verified data and the governing codes (the National Building Code of India, ECBC, the relevant IS standards). Any projection, figure or file this course names is illustrative and scenario-dependent, never a specification and never a prediction of what a particular future year will actually bring.