The Gold Standard of Vacuum Science | Why Two Connected Volumes Can Produce Two Pump-Down Rates

Call the upstream chamber volume V₁ and its pressure p₁. A downstream volume V₂ has pressure p₂ and connects to a pump of speed S. A connection between the two regions has conductance C. There is no direct pump attached to the first chamber in this model.

Assume a common constant temperature, fixed volumes, an ideal-gas inventory relation and uniform pressure within each region. Assume that constant linear conductance describes the connection over the interval being studied. These restrictions define a mathematical model rather than a universal description of evacuation.

Gas stored in the second volume is not automatically removed the instant it crosses the connection. Its inventory can increase or decrease as inflow and pumping compete. That storage is the feature we will retain.

A pressure record therefore needs its measurement location. The two gauges can report different values without either being wrong. Whether their indicated values follow these physical pressures also requires suitable calibration and response time, which we initially idealise away.

Contents · Next chapter

Use litres for both volumes, litres per second for C and S, pascals for pressures, and seconds for time. A throughput then has units Pa L/s. These units are consistent as long as each term uses the same volume convention.

The connecting throughput from region one to region two is C(p₁ − p₂). The pump removes throughput Sp₂ at its inlet. Neither C nor S alone is a fixed number of molecules removed per second; each multiplies the relevant pressure or pressure difference.

At fixed temperature, pressure times volume is proportional to particle population. Consequently V₁ dp₁/dt and V₂ dp₂/dt measure changes in the respective inventories in pressure-volume units per second.

Do a units check before solving anything. A pumping term Sp₁ placed in the first equation would describe a different physical arrangement unless S had already been defined as an appropriate effective speed. We want to derive that approximation, not assume it while retaining another downstream state.

A unit conversion offers another independent check. One litre is 0.001 cubic metres. Converting both volumes and both volume-per-time coefficients from litres to cubic metres leaves the ratios C/V₁, C/V₂ and S/V₂ unchanged. The predicted pressure traces and decay rates must therefore remain the same. Converting the volumes while leaving speeds numerically unchanged would instead create a different system with rates a thousand times too large.

The total inventory changes its numerical unit under that conversion: 2 Pa L becomes 0.002 Pa m³. This is not a change in the gas population. Keeping pressure, storage and throughput units visible makes it possible to compare two calculations without mistaking a different volume convention for a physical disagreement. The model uses volume consistently; it does not depend on choosing litres as a preferred scientific unit.

Contents · Previous chapter · Next chapter

With no incoming gas sources, the balances are V₁ dp₁/dt = −C(p₁ − p₂) and V₂ dp₂/dt = C(p₁ − p₂) − Sp₂.

The connecting flow has opposite signs because gas leaving the first region enters the second. It is not an independent loss from both. The pump term occurs only in the downstream equation.

Add the equations. The internal exchange cancels, leaving d(V₁p₁ + V₂p₂)/dt = −Sp₂. This is an immediate conservation check on the whole model. Gas transferred between volumes does not disappear from their combined inventory.

If either pressure reaches zero while the other remains nonnegative, its derivative cannot push it below zero: at p₁ = 0 the first derivative is Cp₂/V₁, and at p₂ = 0 the second is Cp₁/V₂. This sign check is consistent with nonnegative physical pressures under the declared assumptions.

Contents · Previous chapter · Next chapter

Set V₁ = V₂ = 1 L, C = 1 L/s and S = 1 L/s. The equations become dp₁/dt = −p₁ + p₂ and dp₂/dt = p₁ − 2p₂, with numerical rate coefficients measured per second.

Start both pressures at 1 Pa. The initial connecting pressure difference is zero, so the first chamber initially loses no gas through the connection. Its derivative at t = 0 is zero.

The downstream pump nevertheless removes gas. The second pressure has initial derivative −1 Pa/s. As it falls, a pressure difference develops and gas begins to flow out of the upstream chamber.

This sequence is the concrete mechanism behind an early shoulder in p₁. The chamber does not have to possess a pressure-dependent pump or a new gas source to depart from one simple exponential. A second stored inventory and its initial state are enough in this model.

Contents · Previous chapter · Next chapter

Seek a homogeneous solution proportional to exp(−rt). Substituting into the two equations produces a two-by-two eigenvalue problem. Its determinant condition is r² − 3r + 1 = 0.

The two positive decay rates are rₛ = (3 − √5)/2, approximately 0.381966 s⁻¹, and r𝒇 = (3 + √5)/2, approximately 2.618034 s⁻¹. Their sum is three and product is one.

Their time constants are 1/rₛ ≈ 2.618 s and 1/r𝒇 ≈ 0.382 s. A decay rate and a time constant are reciprocals; do not attach the rate’s numerical value to a time axis as though it were a duration.

The rates belong to the whole connected system. They are not simply the individual pumping speed divided by each volume. The connection couples both states, so a mode generally includes changes in both pressures at the same rate.

Contents · Previous chapter · Next chapter

Write p₁(t) = A exp(−rₛt) + B exp(−r𝒇t). The initial pressure requires A + B = 1. Its zero initial derivative requires rₛA + r𝒇B = 0.

Solving gives A = r𝒇/√5 ≈ 1.170820 and B = −rₛ/√5 ≈ −0.170820. Thus p₁(t) = [r𝒇 exp(−rₛt) − rₛ exp(−r𝒇t)]/√5 Pa.

One modal coefficient is negative. This does not mean negative gas pressure or a negative molecular population. Modes are mathematical components whose sum describes the physical state. Their coefficients need not each be separate physical inventories.

The formula remains positive for nonnegative time. Factor out exp(−rₛt): its remaining numerator is r𝒇 − rₛ exp[−(r𝒇 − rₛ)t], which is at least r𝒇 − rₛ = √5. The negative fast component helps enforce the initial zero slope while the total pressure stays positive.

Contents · Previous chapter · Next chapter

The first balance gives p₂ = p₁ + dp₁/dt for our chosen numerical coefficients. Apply this to the solved upstream expression rather than guessing another exponential.

The result is p₂(t) = [(1 + √5)/(2√5)]exp(−rₛt) + [(√5 − 1)/(2√5)]exp(−r𝒇t). Its coefficients are approximately 0.723607 and 0.276393.

Both sum to one, giving the correct initial downstream pressure. Its initial derivative is −1 Pa/s, as the original balance required. Substitution back into the second equation supplies another check on the solution.

The pressure ratio p₂/p₁ changes during the transient. Treating it as the steady value at every time would remove the very storage behaviour described by the second differential equation. A valid relation at equilibrium is not automatically an identity along every transient trajectory.

Contents · Previous chapter · Next chapter

The familiar steady effective speed is CS/(C + S). In our example it is 0.5 L/s. Applied as an immediate one-volume model to V₁ = 1 L with zero source, it predicts p₁,approx(t) = exp(−0.5t) Pa.

These are rounded values from the stated fictional equations, not apparatus measurements. The approximation predicts a faster reduction of chamber pressure for this starting state.

At late times, the actual upstream decay rate approaches 0.381966 s⁻¹, rather than 0.5 s⁻¹. The discrepancy is not a numerical rounding effect. The two models retain different storage assumptions and therefore have different characteristic equations.

Contents · Previous chapter · Next chapter

Set the downstream storage term V₂ dp₂/dt to zero. Its balance becomes C(p₁ − p₂) = Sp₂, giving p₂ = [C/(C + S)]p₁.

Substitute this relation into the first balance. The chamber then obeys V₁ dp₁/dt = −[CS/(C + S)]p₁. This is the effective-speed formula used dynamically after the downstream state has been eliminated.

The derivation identifies the assumption precisely: downstream accumulation is being neglected. It does not say that the connecting component literally has zero geometric volume or that a real gauge must always read the algebraic ratio.

A small downstream storage time relative to the timescale of upstream change can support this approximation over a useful interval. Its adequacy should be tested against the intended accuracy and initial transient, rather than accepted solely because the steady conductance relation is correct.

Contents · Previous chapter · Next chapter

For arbitrary positive V₁, V₂, C and S, substituting exp(−rt) into the two balances gives V₁V₂r² − [CV₂ + (C + S)V₁]r + CS = 0.

The rate sum is C/V₁ + (C + S)/V₂, and the rate product is CS/(V₁V₂). These expressions provide independent dimensional checks. Both terms in the sum have units s⁻¹, while the product has units s⁻².

The rates are real. Writing a = C/V₁ and b = (C + S)/V₂, the discriminant can be expressed as (a − b)² + 4C²/(V₁V₂), which is positive. Positive sum and positive product then give two positive rates.

These conclusions rely on positive finite parameters and the specified linear connection. A disconnected region, absent pump or changing flow law changes the characteristic problem. The limiting cases are informative, but should not be substituted into an equation after dividing by a parameter that has become zero.

Contents · Previous chapter · Next chapter

Let V₂ become small while V₁, C and S stay fixed. One rate becomes large, approximately (C + S)/V₂. The other approaches CS/[V₁(C + S)], the effective-speed decay rate.

For V₁ = 1 L, C = S = 1 L/s and V₂ = 0.1 L, the rate equation is r² − 21r + 10 = 0. The slow rate is approximately 0.4875 s⁻¹ and the fast rate approximately 20.5125 s⁻¹.

The slow rate is now much closer to the one-volume prediction 0.5 s⁻¹. A short initial adjustment remains possible because two independent starting pressures still need to settle toward the slow relation.

This is why reducing a state can work well after an initial interval but fail at the very beginning. Compare the timescale of interest with the fast adjustment, and state which part of the record the reduced model is intended to represent.

Contents · Previous chapter · Next chapter

Return to equal one-litre volumes and unit conductance and speed. Keep p₁(0) = 1 Pa, but set p₂(0) = 0 instead of one. The initial chamber derivative becomes −1 Pa/s rather than zero.

The upstream solution is now p₁(t) = [(1 + √5)/(2√5)]exp(−rₛt) + [(√5 − 1)/(2√5)]exp(−r𝒇t). The downstream pressure is p₂(t) = [exp(−rₛt) − exp(−r𝒇t)]/√5.

The hardware parameters and decay rates are unchanged. Only the modal amplitudes differ. The chamber begins with a rapid fall because it immediately supplies gas to the initially lower-pressure neighbouring volume.

This comparison warns against interpreting two different early chamber traces as proof of different pump performance. An unrecorded downstream starting state can change the observed transient even when every coefficient in the model is identical.

Contents · Previous chapter · Next chapter

Set p₁(0) = 0 and p₂(0) = 1 Pa, retaining the same parameters and no external gas sources. The first equation gives dp₁/dt at zero equal to +1 Pa/s.

The upstream pressure initially rises because gas moves from the downstream region into it. The downstream pump still removes gas from the combined system. The whole-system derivative remains −Sp₂, so total pressure-volume inventory decreases.

Local accumulation and total accumulation are different quantities. A rising pressure at one location does not by itself prove that gas has entered the whole apparatus from outside. It may reflect redistribution among stored regions.

This is a deliberately simplified counterexample, not a diagnosis of a real pressure rise. Temperature variation, surface release, instrument response and other effects remain possible. The scientific lesson is that a local trace must be interpreted through a stated system boundary before its derivative is labelled an external gas load.

Contents · Previous chapter · Next chapter

Introduce constant throughput Q₁ into the upstream volume and no source into the second. The balances become V₁p₁′ = Q₁ − C(p₁ − p₂) and V₂p₂′ = C(p₁ − p₂) − Sp₂.

At equilibrium, Sp₂,∞ = Q₁ and C(p₁,∞ − p₂,∞) = Q₁. Therefore p₂,∞ = Q₁/S and p₁,∞ = Q₁/S + Q₁/C.

For Q₁ = 0.1 Pa L/s and the unit example parameters, the downstream floor is 0.1 Pa and the upstream floor 0.2 Pa. The same steady effective speed correctly gives Q₁/Seff = 0.2 Pa at the chamber.

This is a useful contrast: the effective-speed relationship can predict the steady chamber pressure exactly under these assumptions while failing to describe the complete transient as one exponential. A model’s success for one observable or limit should not be expanded into a claim about every time-dependent feature.

Contents · Previous chapter · Next chapter

Define deviations a₁ = p₁ − p₁,∞ and a₂ = p₂ − p₂,∞. For constant sources and unchanged parameters, their equations are exactly the homogeneous two-volume balances already solved.

The physical pressures can therefore consist of a nonzero floor plus two decaying modal components. Taking the logarithm of raw pressure does not isolate a single decay mode. Even subtracting the correct floor leaves two components unless one is absent or negligible over the fitted interval.

If the source changes with time, subtracting a single equilibrium no longer removes that forcing. A changing source can generate additional shapes, and the present calculation cannot identify it uniquely from appearance.

For an independent exercise, start the loaded system at its equilibrium pressures. Both derivatives should be zero. Perturb only the chamber pressure and reconstruct the homogeneous deviation. This checks that source terms have been assigned to the correct location and that the floor has not been confused with a gauge offset.

Contents · Previous chapter · Next chapter

In the equal-start no-source example, define an instantaneous upstream decay rate k(t) = −p₁′/p₁ while p₁ is positive. At the start it is zero. At late times it tends to rₛ.

A one-exponential model would have a constant value of k. Our two-mode solution instead produces changing semilogarithmic slope with entirely constant coefficients.

This does not mean every curved trace requires two physical volumes. Several gases, changing surface release, pressure-dependent conductance, pump behaviour and measurement dynamics can also produce departures from one exponential.

The correct conclusion is conditional: a two-volume model is one explicit mechanism capable of generating the observed kind of curvature. To identify it in an actual system, retain geometry, pressure locations, initial conditions and independent information that distinguishes it from competing explanations. A visually appealing two-exponential fit is not enough to count physical compartments.

Contents · Previous chapter · Next chapter

So far, p₁ and p₂ are physical pressures, and measurements are assumed to track them immediately. A real instrument can introduce its own response time or average over an interval, changing the displayed trace.

A simple illustrative gauge response could be written τg R′ + R = p₁, where R is an indicated pressure and τg a response time. That adds another state to the mathematical description. Its transient should not automatically be attributed to another gas volume.

The same issue appears in the opposite direction: fitting a single gauge-lag model to a trace affected by actual gas redistribution can misassign the mechanism. Distinguishing them needs suitable response information or additional measurements, not just another fitting parameter.

The existing gauge-offset and gas-load guide addresses a different measurement ambiguity. Keep offset, dynamic response and physical storage separate in the model even when they can all influence what a recorded curve appears to show.

Contents · Previous chapter · Next chapter

Suppose a chamber trace provides two decay rates accurately. Their sum and product constrain combinations of V₁, V₂, C and S. They do not by themselves give four independently known physical parameters.

The modal amplitudes also depend on the initial pressures and the measurement response. Additional constraints can help, but identifying their number and validity is part of the inference problem.

Even with both pressure records, multiplying V₁, V₂, C and S by the same positive factor leaves both no-source differential equations unchanged after division by their respective volumes. The same pressures versus time can therefore correspond to different absolute volume and throughput scales.

An independent known volume or equivalent scale information can break that particular ambiguity. This example is an exact invariance, not merely a consequence of noisy data. More precise sampling of unchanged traces cannot identify a scale that the mathematical observables do not contain.

Contents · Previous chapter · Next chapter

Integrate the combined no-source balance from zero to t. The result is V₁p₁(t) + V₂p₂(t) = V₁p₁(0) + V₂p₂(0) − S∫₀ᵗp₂(u)du.

The loss of stored gas must equal the integrated removal at the pump within the model. This check uses the whole pressure history and the actual pump-inlet pressure, rather than assigning chamber pressure to the pump term.

For the equal-start unit example, the initial inventory is 2 Pa L. Both pressures tend to zero, so the integral of p₂ over all time is 2 Pa s. Multiplying by unit pumping speed recovers the initial inventory.

For the alternative start p₁ = 1, p₂ = 0, the corresponding integral is 1 Pa s. The different modal amplitudes preserve the different initial inventory. These relationships provide useful independent checks on analytic solutions and numerical integration.

Contents · Previous chapter · Next chapter

Let S = 0 while C remains positive. There is no pumping, so combined inventory is constant. The pressures approach a common value weighted by the two volumes. One rate becomes zero, representing the conserved total, rather than continued decay to an empty model state.

Let C = 0 while S stays positive. The upstream chamber is disconnected and keeps its starting pressure in the no-source model. The downstream volume independently decays at rate S/V₂. Again a zero rate appears.

These cases make the characteristic polynomial’s constant term CS meaningful. With positive connection and pumping, both inventories can eventually be removed. Removing either link leaves a mode that does not decay.

Do not interpret the limiting equations as instructions to alter actual apparatus. They are mathematical tests of the model’s structure. A derived solution that predicts removal from a disconnected, source-free chamber would fail a basic physical consistency check.

Contents · Previous chapter · Next chapter

Keep C and S at 1 L/s, but double both volumes to 2 L. Every derivative for a fixed pressure pair becomes half its earlier value. Therefore the pressure solution at time t is the original solution evaluated at t/2, with the same initial pressures.

The rates become rₛ/2 and r𝒇/2, and the time constants double. The equilibrium pressures under constant source remain Q₁/S and Q₁/S + Q₁/C because steady storage derivatives vanish.

This separates a transient change from a steady-state change. Increasing stored volume can slow the pressure evolution without changing the final pressure under the same gas load and transport coefficients.

For a learner, predict that result before using a solver. Then check a single point: the doubled-volume system at two seconds should match the original system at one second. If the computed final pressure changes solely because both volumes were doubled, inspect whether a source or speed was unintentionally changed too.

Contents · Previous chapter · Next chapter

The equations resemble a two-node linear circuit: pressure corresponds to node voltage, volume to storage capacitance, conductance to a connecting electrical conductance and pumping speed to a conductance toward a zero reference.

The analogy is useful because it preserves the balance structure. Internal exchange appears with opposite signs, node storage creates time derivatives, and a path to the reference removes the total stored quantity.

It does not equate the physical units or mechanisms. Gas transport still depends on a flow regime, species and temperature. A literal electrical component formula cannot be substituted for a vacuum conductance without justification.

P. Chiggiato’s primary lecture, Vacuum Technology for Ion Sources, introduces time-dependent vacuum calculations and the vacuum–electrical analogy. Our two-node coefficients and solved initial states are original teaching constructions. The analogy helps organise them; conservation and the original gas balances remain the final checks.

Contents · Previous chapter · Next chapter

A useful record identifies both physical volumes, the connection, pump-inlet location, gas assumptions and temperature. Preserve synchronized pressure readings, gauge response information, initial conditions and the interval over which constant parameters are being proposed.

Check whether each region can reasonably be represented by one uniform pressure. A long distributed system may require more spatial detail. Adding two lumps does not automatically make every geometry a two-state system.

Compare predictions at more than one location when the available evidence allows it. A model that fits the upstream trace but predicts an incompatible downstream trace has failed a stronger test than fitting the same curve more closely.

Describe unresolved alternatives. Surface sources, changing composition, gauge response and pressure-dependent transmission may remain uncertain. The goal is a traceable explanation whose assumptions can be checked, rather than a declaration that two fitted rates prove two particular volumes are the cause.

Contents · Previous chapter · Next chapter

Begin the repair route with a sketch and arrows. Ask what enters and leaves each volume, then add the balances and cancel the internal exchange. If the connection appears as a whole-system loss, repair that sign error before calculating modes.

For stabilisation, work the initial derivatives and the equal-start solution. Check both initial pressures, both derivatives and the long-time limit. These four checks catch errors that one plotted curve may conceal.

For extension, derive the general characteristic polynomial, test the small-V₂ limit and compare different initial downstream pressures. Then examine which parameter combinations a pressure trace can actually identify.

Finish with a changed context such as heat exchange between two idealised thermal stores with one connection to a reference bath. Retain the balance reasoning, but replace units, physical laws and applicable assumptions. Transfer means rebuilding the model around the new storage and transport mechanisms, not relabelling a vacuum equation without checking it.

Contents · Previous chapter · Next chapter

In the slow mode of the unit example, the first balance gives p₂/p₁ = 1 − rₛ = (√5 − 1)/2, approximately 0.618034. The quasi-steady downstream approximation instead imposes p₂/p₁ = C/(C + S) = 0.5.

Both pressures can therefore decay at the same late rate while retaining a ratio different from the steady-flow ratio. Common semilogarithmic slopes do not establish that the downstream inventory has stopped changing.

At late times, connecting throughput divided by chamber pressure is C(1 − p₂/p₁), approximately 0.381966 L/s. Pump removal divided by chamber pressure is S(p₂/p₁), approximately 0.618034 L/s. These throughputs differ because the downstream volume is also losing stored gas.

The pump removes more than the instantaneous incoming connection flow, with the difference supplied by declining downstream inventory. Treating those two throughputs as equal would erase that term from the balance. This numerical comparison makes the approximation’s physical meaning visible without changing the pump or invoking a time-varying conductance.

Contents · Previous chapter · Next chapter

Choose initial pressures p₁(0) = 1 Pa and p₂(0) = (√5 − 1)/2 Pa. This initial state lies exactly on the slow-mode pressure ratio. Both pressures then follow a single factor exp(−rₛt).

The fast mode has zero amplitude, although the second volume is still one litre and its storage remains in the equations. A system can possess a mode that a particular initial condition does not excite.

Consequently, observing one exponential does not prove that a physical system has only one state. A second mode may be absent from the chosen trajectory, too fast for the observation interval, too small relative to measurement uncertainty or insufficiently visible at the measured location.

This is another distinction between system parameters and an observed record. The rates describe the system, while the amplitudes depend on its preparation and observation. Reproducing a curve is useful evidence, but the count of visibly fitted exponentials is not an automatic inventory of physical compartments.

Contents · Previous chapter · Next chapter

Allow constant sources Q₁ into the upstream chamber and Q₂ into the downstream volume. At equilibrium the connection carries Q₁, while the pump removes Q₁ + Q₂. Thus p₂,∞ = (Q₁ + Q₂)/S and p₁,∞ = (Q₁ + Q₂)/S + Q₁/C.

For unit C and S, pressures p₁ = 0.3 Pa and p₂ = 0.2 Pa correspond to Q₁ = 0.1 Pa L/s and Q₂ = 0.1 Pa L/s. Multiplying chamber pressure by the steady effective speed 0.5 L/s would give 0.15 Pa L/s. That is neither the upstream source nor the total source in this arrangement.

The usual relation p₁ = Q₁/Seff assumed that the downstream source was zero. Its failure here does not invalidate conductance or pumping equations; it identifies a changed source location that the simplified formula omitted.

With known C and S and both suitable steady pressure measurements, the two source terms can be reconstructed in this ideal model. With only one chamber pressure, multiple source allocations can give the same reading. Unknown gauge offsets or uncertain coefficients add further ambiguities. Keep the source boundary and measurement information beside any inferred load, rather than reporting a fitted rate as proof of a particular physical leak.

Contents · Previous chapter · Next chapter

Does a second decay rate prove a leak?

No. The no-source model here has two decay rates with constant coefficients. A pressure curve alone cannot identify a gas source or exclude competing mechanisms.

Is the steady effective-speed formula incorrect?

It is correct for the stated steady relation and can support a useful reduced transient model when downstream accumulation is negligible. Retaining finite downstream storage changes the transient equations and can change their slow rate.

Does a negative modal coefficient imply negative pressure?

No. Modal components are a mathematical decomposition. The original example has one negative coefficient while its total physical pressures remain positive.

Can a rising local pressure coexist with overall evacuation?

Yes. Gas can redistribute into one region while the pump removes gas from the combined system. The integrated whole-system balance distinguishes local accumulation from net external supply.

An independent final check should start from the balances rather than the displayed mode formulas. Evaluate both derivatives at the stated initial pressure pair, verify cancellation of the connecting throughput in the total inventory, and compare the long-time limits with the source assumptions. These checks remain available even when a learner has not yet mastered eigenvalues. They also reveal whether a numerical answer belongs to the intended physical arrangement or merely solves a differently connected set of equations.

Contents · Previous chapter · Next chapter

Return to Vacuum Science for particle inventory, units, flow regimes and the single-volume balance. Inferring a Connection’s Conductance from Two Pressure Readings develops a steady endpoint inference; it should not be applied to a transient without accounting for storage and measurement timing.

For other causes of misleading records, read Separating Gauge Offset from Gas Load, Separating Temperature Change from Gas Accumulation and Residual versus Incoming Gas Fractions. These address measurement response, temperature and species-dependent removal, respectively, rather than the two-volume mode calculation.

The primary technical routes are Chiggiato’s linked CERN lecture and Leybold's conductance explanation. Sources were checked on 25 October 2026. They support the balance and transport framework; all numerical pressures, parameters, mode calculations and transfer checks here are independently constructed examples.

Connect the lesson with Learning That Transfers into Life and the How X Works hub. A useful explanation identifies where gas is stored, where it moves, where it leaves and which observation would distinguish a physical mode from an alternative explanation.

Contents · Previous chapter · Return to Vacuum Science