space-rha.comShort courseChapter 4 · Mechanisms and technologies

Chapter 4 · Lesson 1 of 6 · about 24 minutes

From a track to a consequence

An ion leaves a short trail of electron–hole pairs in the device. Some of that charge reaches a sensitive node. Depending on the circuit, a bit may change, an output may pulse, a function may stop or a power path may fail. Start the assessment from the function that is affected and how it recovers, and then work back to the physical response that produced it.

Ion track and charge collection

The homepage animation of space-rha.com: an ion crosses a CMOS inverter drain, generates electron-hole pairs, and the collected charge becomes a current pulse at the node.

An original space-rha.com laboratory, embedded unchanged with its own controls; its onset values, gains and depths are teaching assumptions, not limits for flight hardware.

The circuit decides what a strike becomes.

Generated, collected and critical charge are three different quantities.

LET gives the energy lost per unit mass thickness along the track. It says nothing about how much of the resulting charge reaches any one node. The junction field, where the track passes, diffusion, the well structure, the bias and the neighboring nodes all affect how much is collected. Whatever response model you apply has to match the state the circuit was in and what the test could observe.

The same deposited energy can therefore deliver different charge to the node. Whether that charge upsets anything depends on the critical charge of the node, which the circuit sets.

Evidence to retain. The record should name the technology, the supply voltage, the device state, the ion energy and range, the angle of incidence, which functions were monitored and what counted as recovery, because each of these limits where the result can be applied. TCAD answers questions about collection; circuit simulation answers questions about the response to the collected charge.

Mission context. A computer on the ISS sees trapped protons, mostly in the SAA, together with galactic cosmic rays and whatever solar particles reach that orbit. Getting from those species to a device response has to go through nuclear secondaries and, where it applies, direct proton ionization.

One particle can upset more than one cell.

A memory upset is not by itself a mission failure. What it costs depends on what was stored, on the error correction, on the interleaving and scrub interval, and on whether one strike can defeat the code. A physical multiple-cell upset and multiple errors in one logical word are different things, so count them separately.

ECC corrects an error pattern, not a particle count.

Charge sharing and the well response can disturb several storage nodes at once. A code that corrects one bit per word can still recover a physical cluster if the layout interleaves those bits into different words. The reverse also happens: an error in a shared peripheral or in the controller can get past array-level ECC altogether.

The test record should keep the raw addresses, the timestamps and the grouping of errors into events, since those are what let you tell a cluster from a coincidence. The code, the bit mapping and the scrub behavior should be the ones the flight design uses, and the flux should be low enough that each error can still be attributed to one event.

Evidence to retain. Keep the addresses and error patterns both before and after correction, along with any functional interrupts, the scrub latency and whatever was lost in recovery. Report the event cross section and the bit-error cross section as separate numbers.

Mission context. A long-duration recorder can tolerate isolated corrected bits and still lose data the first time one cluster exceeds its code or stops its controller.

A transient matters when something captures it.

A pulse from a particle strike can die out in the logic, be attenuated on the way, or arrive between two sampling edges and go unnoticed. It can also propagate to a latch and become stored state. Whether it matters depends on the clocking, the logic path, what the output drives and what a false action would cost.

Three masking mechanisms, each with its own evidence.

Logical masking means the path was not sensitized. Electrical masking means the pulse lost amplitude or width on the way. Timing masking means the pulse missed the capture window. Clock frequency alone does not set the susceptibility, because path delay, pulse broadening and the behavior of the sequential cells change with it.

The test should exercise the states, frequencies and paths the application uses, because each of the three masking mechanisms depends on them. Where the application analysis needs them, measure the distributions of pulse width and amplitude rather than only counting events.

Evidence to retain. Capture transients with enough bandwidth, or with event counters that have been validated for the pulses in question, and state the dead time, the pulse thresholds and the workload that was running. Electrical injection results are worth comparing against what the beam produced.

Mission context. An asynchronous command line can react to a pulse that would be harmless on a sampled telemetry bus.

A short pulse can leave a long disturbance in the system.

An analog SET can saturate an amplifier, trip a comparator, corrupt a conversion or knock a clock loop out of lock. The duration of the pulse at the struck transistor is generally not the recovery time at the system output. The application has to say what amplitude, duration or sequence of errors it can tolerate before a test endpoint can be chosen.

Load and operating point change the response.

Gain, bandwidth, feedback, rail headroom and output load all shape an analog transient. An ADC can show isolated bad samples, bursts of them, or a functional interrupt. A PLL can show a phase disturbance or lose lock. No single output-voltage threshold covers all of these endpoints.

The test should sweep the bias, load, input state, sample rate and temperature that the application spans, and it should record the waveforms, including saturation and recovery, because an event count alone does not say what the circuit downstream would have seen.

Evidence to retain. Record the load, the input waveform, the sample and clock state, the detection bandwidth and the recovery criterion. Capture the transient waveforms and also any state change that persists after the transient is gone.

Mission context. A false overcurrent trip can interrupt a payload even when the regulator itself returns to normal in microseconds.

The particle that arrives may not make the damaging track.

Protons and neutrons can produce energetic recoils and reaction products in the device and in the materials around it, and those charged products deposit energy near sensitive nodes. In some small or sensitive technologies, protons can also cause single-event effects directly through their own ionization.

Energy and material matter beyond a single LET value.

The proton or neutron response depends on the particle energy, the reaction channels open at that energy, the materials near the sensitive volume and the structure of the sensitive volume itself. High-Z interconnect and packaging can change what secondaries are produced. A heavy-ion screen therefore does not by itself establish the proton response for every endpoint.

Where the data exist, use proton and neutron cross sections as functions of energy. Shielding changes the field that reaches the part, and low-energy protons may or may not reach the sensitive region, so the analysis has to include both.

Evidence to retain. Keep the proton or neutron energy distribution, any degrader material, the TID and displacement damage accumulated during the test, the fluence, the operating modes and the endpoint. Where the data allow it, say whether the response is attributed to direct ionization or to nuclear reactions.

Mission context. For electronics in LEO, the SAA protons can dominate some responses even though the heavy-ion background is still there and still produces events.

A technology label does not settle susceptibility.

Smaller nodes, SOI isolation and new transistor geometries change both the charge collection and the sensitivity of the circuit. A part can improve for one endpoint and become more vulnerable through another. The question to put to any data set is which structure, which state and which function it covers.

Angle is more than an effective-LET correction.

The inverse-cosine rule describes a uniform track through an ideal planar layer. Real sensitive volumes are finite, tracks can enter through a sidewall or cross a well boundary, several nodes can lie along one path, tiers can be stacked, and the shielding differs by direction. All of these change the path and the collected charge. Tilt and azimuth are therefore two separate test variables.

At each orientation, check that the ion still penetrates the package and has range enough to reach the sensitive layer. A front-side, normal-incidence result does not necessarily bound a backside or oblique exposure.

Evidence to retain. When the geometry could change the conclusion, use a model that includes the layout and add angle and energy runs aimed at the question. Record the tilt, the azimuth, the entry side and whether the beam reached the sensitive layer.

Mission context. An advanced processor whose individual cells are each quite tolerant can still need a strong assurance case, because a long mission, a large and complex die and critical control functions all raise the consequence of the events that do occur.

Going deeper

Transport connects the track to a terminal current.

The Poisson and carrier continuity equations couple the electric field to drift, diffusion and recombination. A dense ion track can perturb the field and extend the collection through transient funneling or parasitic bipolar action. The collected charge is the integral of the terminal current, and a static depleted volume alone can miss these processes.

Jn = qμn nE + qDn∇n; Jp = qμp pE − qDp∇p. A Qcrit estimate from the node capacitance alone omits the circuit feedback and the pulse shape.

An SRAM upset is a trajectory through a bistable circuit.

Cross-coupled inverters pull a small perturbation back, but a large enough current injection drives them through to the opposite state. The pulse duration, the node capacitance, the restoring drive and the coupling between nodes all enter. A TCAD-to-SPICE flow that includes the layout can capture charge sharing that a single independent rectangular sensitive volume cannot.

A static Qcrit comparison is a useful estimate. The actual upset boundary is dynamic and depends on the injection waveform, the cell state, the supply and the coupled nodes.

The current waveform meets a nonlinear transfer function.

Charge collected at a drain perturbs the node, and the transistor drive and the load shape the resulting voltage pulse. The gates downstream can narrow it, quench it or broaden it. Whether it is captured depends on the setup and hold behavior, on metastability and on the sequential element itself. Injection at several nodes, or reconvergent paths, can defeat the assumption that each pulse travels alone.

The double-exponential current source delivers the charge it was told to deliver. It is a stimulus for the circuit model, not a self-consistent solution of the ion track or a calibrated model of the logic cell.

Internal nodes can dominate a macroscopic recovery.

Charge collected at a high-impedance node can change a stored charge, a bias current or the state of a comparator. Feedback and large-signal saturation can then stretch the recovery far beyond the collection transient. In a converter, which stage was struck and where the pipeline was at that moment determine which outputs are corrupted.

In a circuit model, inject a physically justified current at the internal nodes and then check the predicted endpoint against the irradiation result. Injection at an external pin may bypass the structure that is actually vulnerable.

Fold in the distribution of local depositions, not an average recoil.

Nuclear reaction products come in a range of species, energies, directions and ranges, and where each one deposits its energy determines how it couples to the collection and the circuit response. Direct proton ionization matters once the local deposition is enough to disturb the circuit, and slowing in the overlying material and the angle of the track can change whether it is.

Coupled nuclear transport with a calibrated device response is far more work than a single LET threshold. The uncertainties in the physics list, the geometry and the cross sections should stay visible in the result.

Collection geometry and circuit topology evolve together.

Isolation can remove one parasitic path while floating-body or bipolar effects remain in particular structures. FinFET and gate-all-around devices change the electrostatics, but shared diffusions, wells, contacts and 3D integration still couple nodes to each other. A rectangular-parallelepiped sensitive volume is an abstraction, and it is valid only inside the domain where it has been checked.

The exact chord of a ray through a finite volume differs from the infinite-slab approximation, and even the exact chord leaves out the collection weighting, diffusion, the track structure and the circuit feedback.

The common mistake

The silicon conversion shown in the laboratory assumes a uniform LET along an effective path. It gives neither a measured cross section nor a general upset threshold.

Takeaway

  • An LET number is not a risk assessment. The analysis has to follow the local field and the circuit response all the way to the consequence for the application.
  • Match the protection to the error pattern the part actually produces and to the way bits map into logical words.
  • Test the timing behavior and the logic states the application uses, since a powered device with a heartbeat exercises very little.
  • Work out what transient the receiver or the control loop can tolerate before choosing the test endpoint.
  • Choose the particles and energies to cover the physical route to the endpoint that was observed.
  • Technology, layout, orientation, and ion energy and range all limit where a result applies.

Check yourself

A colleague multiplies the LET of the worst-case ion by the depth of the sensitive region, compares the result with the critical charge of a cell and declares the cell safe. What is missing from that argument?

Generated, collected and critical charge are three different quantities. LET gives the energy lost per unit mass thickness along the track and says nothing about how much of the resulting charge reaches the node; the junction field, where the track passes, diffusion, the well structure, the bias and the neighboring nodes all affect the collection. The critical charge is set by the circuit, and a static estimate from the node capacitance alone omits the circuit feedback and the pulse shape, so the actual upset boundary is dynamic and depends on the injection waveform, the cell state and the supply.

The course asks a memory test to keep the raw addresses and timestamps, to group the errors into events, and to report the event cross section and the bit-error cross section as two separate numbers. Why?

Because ECC corrects an error pattern, not a particle count. A physical multiple-cell upset and multiple errors in one logical word are different things: a code that corrects one bit per word can still recover a physical cluster if the layout interleaves those bits into different words, and an error in a shared peripheral or in the controller can get past array-level ECC altogether. The addresses, the timestamps and the grouping are what let you tell a cluster from a coincidence, and the flux has to be low enough that each error can still be attributed to one event. The two cross sections feed different parts of the protection design (lesson 10.1).

A digital part is irradiated powered, with a heartbeat running at one clock frequency, and shows few errors. The designer concludes that it will do as well at the flight frequency and workload. What does that overlook?

Whether a transient matters depends on logical, electrical and timing masking, and each of the three depends on the states, paths and frequencies that were exercised. Clock frequency alone does not set the susceptibility, because path delay, pulse broadening and the behavior of the sequential cells change with it, and a powered device with a heartbeat exercises very little of the logic. The test should run the states, frequencies and paths the application uses, and where the application analysis needs them, measure the distributions of pulse width and amplitude rather than only counting events.

A transient test on an analog part counts output excursions above one voltage threshold and reports a cross section. Why can that number be insufficient for a converter, a PLL or a regulator that feeds an overcurrent detector?

No single output-voltage threshold covers all the endpoints an analog transient can produce: an ADC can show isolated bad samples, bursts of them or a functional interrupt, a PLL can lose lock, and a false overcurrent trip can interrupt a payload even when the regulator itself returns to normal in microseconds. The duration of the pulse at the struck transistor is generally not the recovery time at the system output, because feedback and large-signal saturation can stretch the recovery well beyond the collection transient. The application has to say what amplitude, duration or sequence of errors it can tolerate before the endpoint is chosen, and the test should sweep the bias, load, input state, sample rate and temperature and record the waveforms, including the saturation and the recovery.

A part passed a heavy-ion screen at normal incidence from the front side, and the mission is in LEO, where the SAA protons dominate the flux. A reviewer treats both the angular question and the proton question as closed. What is wrong?

They are two separate gaps. The inverse-cosine rule describes a uniform track through an ideal planar layer, but real sensitive volumes are finite, tracks can enter through a sidewall or cross a well boundary, and the shielding differs by direction, so tilt and azimuth are separate test variables and a front-side, normal-incidence result does not necessarily bound an oblique or backside exposure. Protons act through nuclear secondaries and, in some small technologies, through direct ionization, so the response depends on the proton energy, the reaction channels and the materials near the sensitive volume, and a heavy-ion screen does not by itself establish the proton response for every endpoint; where the data exist, use proton cross sections as functions of energy, with the shielding included.

Further reading