Lab safety, ethics and reproducibility

An experiment that injures someone, or produces a result that cannot be trusted, has failed whatever its data say. This lesson covers the two kinds of responsibility that come with experimental work. The first is safety: physics laboratories contain high voltages, lasers, radioactive sources, cryogenic liquids, vacuum systems and heavy equipment, each with well understood hazards and well established ways of controlling them. The second is integrity: reporting what was actually done and found, selecting and analyzing data by fair rules, and making results reproducible so that others can check them. Both rest on the same habits of planning ahead, keeping records and being honest about uncertainty.

Thinking about hazards

Safety in a laboratory starts with a hazard assessment before any work begins: identify what could cause harm, estimate how likely and how serious it is, and decide what controls reduce it to an acceptable level. Controls are ranked in a hierarchy, from most to least effective. Eliminate the hazard if possible, for example by using a low-voltage supply instead of a high-voltage one. Substitute a less hazardous option, such as a lower-power laser. Use engineering controls, such as interlocks, enclosures and shielding, that protect people without relying on their behavior. Use administrative controls: procedures, training, signs and supervision. Last, use personal protective equipment such as safety glasses and gloves, which matters but only works if worn correctly and every time. In the United States, the Occupational Safety and Health Administration's Hazard Communication Standard requires labels and safety data sheets for chemicals, and OSHA's general duty clause requires employers to protect workers from recognized hazards. Universities add their own training requirements, and you must complete them before working in a research laboratory.

Electrical safety

Current through the body, not voltage as such, causes harm. A current of about 1 mA at mains frequency is felt; 10 to 20 mA can cause muscles to contract so that a person cannot let go; currents of order 100 mA through the chest can cause ventricular fibrillation, which is often fatal. Dry skin has a resistance of perhaps 100 kΩ, but wet or broken skin far less, so voltages above about 50 V must be treated as dangerous. Mains wiring, power supplies for photomultipliers and particle detectors, which can exceed 1 kV, and the capacitor banks of pulsed lasers and magnets are the main risks in a physics laboratory. Capacitors deserve special mention because they can hold a lethal charge long after the equipment is switched off: the energy \(CV^2/2\) of a 100 μF capacitor at 1 kV is 50 J. Good practice: switch off and unplug before working on a circuit; discharge capacitors through a suitable resistor and check with a meter; keep one hand away from the circuit when probing high voltage; never defeat an interlock; and know where the emergency power cut-off is. Ground-fault circuit interrupters, which trip when a few milliamperes leak to ground, protect against the commonest shock hazard near water.

Lasers

Lasers concentrate light into a narrow beam, and the eye focuses it further, so even a modest laser can burn the retina instantly and permanently. Lasers are classified by the hazard they present. Class 1 is safe under normal use, often because the beam is enclosed. Class 2 visible lasers below 1 mW are protected by the blink reflex for brief exposure, but staring into the beam is unsafe. Class 3R, up to 5 mW, is low risk but can damage the eye if viewed through optics. Class 3B, up to 500 mW, is hazardous to the eye from the direct beam and from mirror-like reflections. Class 4, above 500 mW, is hazardous to eyes and skin even from diffuse reflections and can start fires. Many research lasers are Class 4, and infrared lasers are especially dangerous because the beam is invisible and the blink reflex does not respond. Controls include enclosing beam paths, keeping beams below or above eye level, removing watches and rings that could reflect a beam, using beam blocks and interlocked doors, and wearing goggles rated for the specific wavelength and optical density. Alignment is the most dangerous time and should be done at the lowest practical power.

Ionizing radiation

Teaching laboratories use small sealed radioactive check sources, typically a few to some tens of kilobecquerels, for counting experiments, half-life measurements and calibration. Their hazard is small if they are handled properly, and the principles that control it scale up to any radiation work: time, distance and shielding. Dose is proportional to time spent near a source. For a point source, the dose rate falls as the inverse square of distance, so moving from 1 cm to 10 cm reduces it a hundredfold. Shielding attenuates radiation: a sheet of paper stops alpha particles, a few millimeters of plastic or aluminum stops beta particles, and lead or concrete attenuates gamma rays. Sealed sources must not be opened, and they should be handled with tongs rather than fingers, kept in their shielded container when not in use, and returned and logged. Radiation work in research is regulated, in the United States by the Nuclear Regulatory Commission or an agreement state, which license sources above exempt quantities, and institutions appoint a radiation safety officer. The aim is to keep doses as low as reasonably achievable, a principle known by its initials as ALARA.

Cryogens, pressure and vacuum

Liquid nitrogen at 77 K and liquid helium at 4.2 K are used to cool detectors, superconducting magnets and samples. They cause severe cold burns on contact, so wear insulated gloves that can be shaken off quickly, a face shield when pouring, and closed shoes. Two less obvious hazards matter more. Cryogens expand enormously when they boil: a liter of liquid nitrogen makes nearly 700 liters of gas at room temperature, so a sealed container can explode, and dewars must always be vented. And the gas displaces air. In a small, poorly ventilated room, boiling off enough liquid nitrogen can lower the oxygen concentration below the level at which people lose consciousness, without warning, since nitrogen has no smell. Oxygen monitors with alarms are fitted where large amounts of cryogens are used, and dewars should not travel in elevators with people. Vacuum systems carry an implosion risk for glass vessels, which should be shielded or taped; compressed gas cylinders must be chained upright and moved with their caps on, because a cylinder whose valve breaks off becomes a projectile.

import math, random
random.seed(397)
# 1. A small teaching check source: the dose rate falls with the inverse square of distance.
#    Cs-137 gives about 0.09 microsievert per hour at 1 m per MBq, so a 37 kBq (1 microcurie) source gives about 0.0034.
G = 0.093*0.037   # microsievert per hour at 1 m for 37 kBq
for r_cm in (1, 5, 10, 30, 100):
    print(f"at {r_cm:3d} cm: about {G/(r_cm/100)**2:9.4f} microsievert per hour")
print(f"handling it for 10 minutes at 30 cm: {G/(0.3)**2*10/60:.4f} microsievert (natural background is about 0.05 to 0.3 per hour)")
# 2. Liquid nitrogen boils into a large volume of gas: why it must never be sealed and why small rooms need ventilation.
rho_liq, M = 808.0, 0.028          # kg/m^3, kg/mol
n_mol = rho_liq*1.0e-3/M            # moles in 1 L of liquid
Vgas = n_mol*8.314*293/101325       # m^3 at 20 C and 1 atm
print(f"1 L of liquid nitrogen makes {Vgas*1000:.0f} L of gas at room temperature")
room = 3.0*3.0*2.5                  # m^3
for vol in (5, 25, 50):
    frac = 0.209*room/(room + vol*Vgas)
    print(f"  {vol:2d} L boiled off in a sealed 3 x 3 x 2.5 m room: oxygen falls to {frac*100:.1f} (out of 100) by volume")
# 3. Multiple comparisons: test 20 independent null effects at p < 0.05; how often is at least one 'significant'?
trials, hits = 20000, 0
for _ in range(trials):
    if any(random.random() < 0.05 for _ in range(20)): hits += 1
print(f"chance of at least one false positive in 20 tests: simulated {hits/trials:.3f}, exact {1 - 0.95**20:.3f}")
print(f"Bonferroni threshold for 20 tests: p < {0.05/20:.4f}")

at   1 cm: about   34.4100 microsievert per hour
at   5 cm: about    1.3764 microsievert per hour
at  10 cm: about    0.3441 microsievert per hour
at  30 cm: about    0.0382 microsievert per hour
at 100 cm: about    0.0034 microsievert per hour
handling it for 10 minutes at 30 cm: 0.0064 microsievert (natural background is about 0.05 to 0.3 per hour)
1 L of liquid nitrogen makes 694 L of gas at room temperature
   5 L boiled off in a sealed 3 x 3 x 2.5 m room: oxygen falls to 18.1 (out of 100) by volume
  25 L boiled off in a sealed 3 x 3 x 2.5 m room: oxygen falls to 11.8 (out of 100) by volume
  50 L boiled off in a sealed 3 x 3 x 2.5 m room: oxygen falls to 8.2 (out of 100) by volume
chance of at least one false positive in 20 tests: simulated 0.636, exact 0.642
Bonferroni threshold for 20 tests: p < 0.0025

The numbers make the principles concrete. A 37 kBq cesium-137 check source gives about 0.003 microsievert per hour at 1 m, far below the natural background of roughly 0.05 to 0.3 per hour, but about 34 per hour in contact at 1 cm: distance is the main protection, and tongs work. Handling the source for ten minutes at 30 cm gives a dose of a few thousandths of a microsievert. The cryogen calculation shows why ventilation matters: boiling 50 liters of liquid nitrogen in a sealed room the size of a small office would lower the oxygen to about 8 of every 100 molecules of air, a level at which people collapse within minutes. The final part of the program turns to integrity, and needs some explanation.

Integrity: honest data and honest analysis

Scientific misconduct is usually defined as fabrication, inventing data; falsification, altering or selectively omitting data or results; and plagiarism, presenting others' work as one's own. Outright fabrication is rare but has happened at high levels: in 2002, a series of spectacular results on molecular-scale transistors and superconductivity from Bell Labs was found to rest on data that had been reused and invented, and dozens of papers were retracted. The commoner problems are subtler and often unintentional. Data are rejected because they look wrong, without a criterion fixed beforehand. An analysis is adjusted, a cut moved, a background redefined, until the result looks significant. Many comparisons are made and only the striking one is reported. Each of these inflates the apparent significance of a result.

The last program in the code above shows why multiple comparisons matter. If you test 20 independent effects that are all truly absent, using the common threshold of p < 0.05, the chance that at least one appears significant is \(1 - 0.95^{20} = 0.64\). Report only that one and you have published a false discovery, without any dishonest intent. The remedies are to decide in advance which tests you will do, to correct the threshold for the number of tests, for example by the Bonferroni rule of dividing it by the number of tests, and to report everything you tested. Particle physicists require a five-sigma result, a chance probability of about 3 in 10 million, before claiming a discovery, partly because they search for effects in many places at once; the Higgs boson announcement at CERN in 2012 met that standard in two independent experiments. Richard Feynman, in his 1974 commencement address at Caltech that became known as Cargo Cult Science, put the principle simply: the first principle is that you must not fool yourself, and you are the easiest person to fool.

Reproducibility

A result becomes knowledge when others can reproduce it. In the 2010s, large efforts to repeat published studies in psychology, medicine and other fields found that a substantial fraction could not be replicated, a problem widely called the replication crisis. Physics has fared comparatively well, partly because its results are often quantitative predictions tested by independent groups, but it is not immune, as the faster-than-light neutrinos and the inflation signal described in lesson 1 showed. Practices that make work reproducible are simple to state: a detailed notebook; raw data preserved; analysis code kept with the data, under version control, so every figure can be regenerated; methods described fully enough for someone else to repeat them; uncertainties and their derivation reported; and data and code shared where possible. Preregistration, publicly stating an analysis plan before collecting data, and blind analysis, described in lesson 2, protect against unconscious bias. Independent replication, ideally with a different method, remains the strongest test.

Authorship and credit

Modern experiments are collaborative, from a pair of students at a bench to thousands of physicists on a detector at CERN. Authors of a paper should have made a real intellectual contribution and should take responsibility for its content; people who helped in other ways are acknowledged. Students should discuss authorship with their supervisors early. Credit the work of others with citations, including the instruments, software and data sets you used. Treat colleagues fairly: harassment and exclusion damage both people and science, and most institutions have codes of conduct and offices that handle complaints.